[This article first appeared in the journal Music Performance Research (MPR) 6: 97-125, 2013. This new version is dated 9 June 2022. The present version was made because the original version as published by MPR had become dysfunctional, as a result of the audio files used in the article no longer being linked to to it and thus being unplayable. The journal's current website address is http://musicperformanceresearch.org/volume-6/. My message of 12 March 2022 to the journal's email address mpronline@rcs.ac.uk was answered on 24 May 2022 with an apology and finishing with "More soon, I hope.". I have not received anything further (by 9 June 2022). Copyright presumably remains with MPR (2013) but because their pdf release is no longer functional the present version seems clearly justified. This version is an html emulation of the original pdf version, with just a few minor layout details being changed. The audio examples are playable here (9 June 2022). Several documentary updates are enclosed in {braces}.]

The reconstitution of historical piano recordings: Vladimir de Pachmann plays Chopin's Nocturne in E Minor

by Nigel Nettheim

The MARCS Institute, University of Western Sydney, Australia {now named Western Sydney University}

ABSTRACT: The reconstitution of historical music recordings from close measurement of their details has seldom been attempted. Three different recordings of the same piece by the pianist Vladimir de Pachmann, one on gramophone and two on rolls, are here reconstituted. The timing, loudness and pedalling on the gramophone recording are measured by listening to slow playback, since no adequate algorithmic or automated method exists. The timing on the rolls is measured by a mechanical reading device; because loudness and pedalling are considered inadequately specified on rolls, those measurements are taken from the gramophone recording as a first approximation, to be refined when necessary. All measurements are converted to midi values for playback, realized here on a virtual piano. For general (non-specialist) listeners, the resulting reconstituted versions of all three performances seem of high quality, and may therefore have value independently of this article: the gramophone recording acquires a modern sound, and the rolls approach their full potential. For music students, analysis of the performances is facilitated; in particular, the reconstitutions may be separated into single musical voices to be played back alone, providing an instructive and otherwise unobtainable resource.

KEY WORDS: Historical piano recordings, reconstitution, melody, gramophone recordings, piano rolls, Vladimir de Pachmann.

The importance of historical music recordings is widely acknowledged; historical piano recordings in particular will be dealt with here. Their reconstitution may provide benefits of two kinds: improved sound quality for general listeners, and clearer revealing of the details of the performance for those who wish to study them.

My work began with the reconstitution of a gramophone recording. Here the benefit of


the first kind may be debatable, for some general listeners might prefer a conventional transfer of a 78 rpm recording with its somewhat nostalgic-sounding 'historical haze', [1] together with greater authenticity than a reconstitution can guarantee. Also the modern virtual piano used here, though usually considered to be of high standard, might not be to the taste of all listeners. The benefit of the second kind is more certain for, despite the limitations of any reconstitution, that is how some features of the playing can best be studied. A prime example is the opportunity to extract from a gramophone recording musical voices, such as a melodic line and a bass line, and to play the voices back separately; this can reveal a performer's approach to the treatment of the strands of the music, and could not have been achieved otherwise than by reconstitution. I have, then, reconstituted a gramophone performance as accurately as I could, not changing the original performance in any arbitrary way such as by correcting notes or imposing my own ideas of appropriate loudness, timing, articulation or pedal. The only change I considered desirable was to avoid the loss of loudness in high and low notes brought about by the original recording process. [2] The reconstitution and the source can readily be compared.

The only past attempt known to me with an aim at all similar to the present one is the commercial work of Zenph Studios (2007, 2009). [3] We read that "Zenph's team has developed a process for working backwards, from an original audio recording to a detailed description of how every note was played" (Zenph Studios, 2012). Zenph's process remains proprietary. [4] A process that may be worth mentioning in this connection is that of approximating the reconstituted wave-form to the source wave-form with iteratively modified parameter values (that is, descriptive details), the estimated spectral densities of the waveforms being available to assist the approximating. That process would suffer from at least two drawbacks: (i) the process might not converge to the target in all cases because the slope of the high-dimensional distance function would be unknown and a local rather than a global minimum might be reached, and (ii) different sound qualities of source and target strongly interfere with wave-form comparison. Least of all could one expect to determine refined pedalling by this method.

Zenph's process, whether it was the one just mentioned or another, is in any case inadequate for artistic purposes. This has been indicated by Malik (2011), who compared source and reconstitution (Zenph's 're-performance') in respect of loudness patterns, balance of chords (the relative loudness of their constituent notes), pedalling and other vital features which are particular focuses of the present work. I have confirmed Malik's observations. Finally, Zenph's measurements, however they were arrived at, would need to be modified for the current condition of the particular physical piano used for playback, a possible further source of error, by comparison with the stable virtual instrument used here.

[1] By 'historical haze' I mean features such as noise, frequency response limitation (reduced loudness for high and low notes), and dynamic range compression (restricted range of loudness for all notes).

[2] This frequency response limitation could in any case be remedied with an equalizer when playing a gramophone record or sound file.

[3] Slåttebrekk (2010) sought to reproduce in his own playing the kinds of nuances found in Grieg's 1903 recordings, a considerably different aim from the present one.

[4] Zenph's patent applications can be inspected at http://www.faqs.org/patents/inventor/walker-ii-us-5/, retrieved 2 February, 2013.


Following the gramophone reconstitution, my work continued with the reconstitution of two reproducing piano roll performances (the three original performances are not the same – they were recorded at different times). Piano rolls encode the timing of a performance well, but beyond that they have considerable limitations, varying somewhat with the particular manufacturer's proprietary system. For my work, it is the heard limitations that are relevant, rather than the technical explanation for them. It may nevertheless be observed that the fine control of loudness especially at soft levels is limited, the loudness of individual notes within a chord could usually not be indicated, pedalling indications recorded during a performance were of unknown reliability, the degree of partial pedalling was not recorded at all, and the quality of the editor's contribution can always be questioned. I sought to remedy those limitations in ways to be described, including reference to details already measured for the somewhat similar gramophone performance. The sound quality is easily controlled with a virtual instrument, so dependence on the condition of a physical reproducing piano is removed. No previous attempt to reconstitute a roll performance with replacement of its missing information is known to me.

For both recording media, gramophone and rolls, I hope to have provided to a fair extent the first kind of benefit, improved sound quality, but it is the second, the instructive study of details, which is the main focus of the present paper. The technical method employed for reconstitution will be dealt with in as much detail as space allows. Possibilities for mutual benefit between the study of gramophone and roll recordings will also be pointed out.



Vladimir de Pachmann (1848-1933) was a famous and very individual pianist, well suited to this project. There are four possible sources for any piece played by Pachmann: a gramophone recording, Welte roll, Duo-Art roll, and annotated edition (Nettheim, 2000, 2001b). I chose the Nocturne in E minor, op.72/1, for which the first three sources exist. The sound quality of the gramophone recording of this piece is relatively good, and I preferred not yet to take on the additional challenge of working with poor sound.

Sound recordings

Pachmann recorded the present piece on 78 rpm discs by the electrical process five times, but the only take that survives is the second one (Chopin, 1827/1927); I chose to refer to the Dante CD (Chopin, 1827/1996) (Audio 1). Pachmann also recorded the piece for Welte Licensee piano roll (Chopin, 1827/1906 or 1925) and for Duo-Art piano roll (Chopin, 1827/1923 or earlier). The Duo-Art roll exists as a London release (016) and as a New York release (6675) with possibly different editing; I chose to use the latter.

Audio 1:         (Or click here for an alternative playing method, then the browser's back-arrow.)

Score editions

Performing artists may consult autographs and manuscripts but they may nevertheless rely largely upon a published edition. For the present recordings it was possible to show that the latter method had been used. The present Nocturne has been published in many editions, including a number available in Pachmann's time (Grabowski, 2009). I consulted nine of


those early editions for comparison with Pachmann's playing. [5] For five key spots in the piece, [6] Pachmann's playing in all three performances matched one score only, yielding the conclusion that he used the Klindworth edition published by Augener (1882). Further, Pachmann's widow published an edition of some of Chopin's works based on his own copies of the score (Chopin, 1934, 1935, 1937), and that too was published by Augener, suggesting confirmation of my conclusion. As if to echo that conclusion still further, Figure 1 shows a postcard of Pachmann indicating copyright by Augener. [7] For close study, the reader would therefore benefit from (internet) access to the Klindworth (Augener) edition. [8]

Postcard of Pachmann

Figure 1. Postcard with Pachmann (recto) and Augener copyright notice (verso).

[5] Casella (c.1915), Debussy (1915), Joseffy (1915)*, Klindworth/Scharwenka (referred to here as Klindworth) (1882)*, Kullak (1881)*, Mikuli (1894)*, Pugno (1902), Sauer (1920), Scholtz (1879)*. Here * indicates retrieval 1 October, 2012, from http://imslp.org/ .

[6] Bar 14, two notes D5; bar 43 chord A4 and F sharp 3 repeated; bar 46 elaborate chord with low E1 (the most distinctive identifier); bar 56 upward stem and slur indicating highlighting; bar 57 B3 present.

[7] This postcard, and a companion one of Augener's, may be corroborative but they do not by themselves provide sufficient evidence, for another postcard with a different photo shows "Breitkopf & Härtel, 54 Great Marlborough Street London W", without a copyright sign.

[8] I confirmed my conclusion with the Paderewski (1951)* edition which, although widely used today, should generally not be used in connection with earlier performances such as the present ones.


Sound output

Before performance details are measured, a particular output method should be chosen, for loudness and pedalling values may have different effects on different instruments. I chose the Pianoteq virtual piano, a software implementation of a mathematical model of the piano's sound. [9] I wrote a computer program in the Matlab language to convert my measurements to a midi file according to the specifications required for playing the music on that output device. Pianoteq provides a number of virtual instruments, from which I chose the 'D4 Classical BA' one based on a Steinway model D from Hamburg ('BA' refers to the microphone placement). Each Pianoteq instrument may be modified through many parameters grouped in four categories: tuning, voicing, design and output; my modifications were mainly in voicing, where I softened the hammers and reduced the hammer noise. The data for the Matlab program is a matrix, each row containing the details for one musical event, whether a note or a pedal action. The details for the notes include a voice number which I assigned to distinguish melody, bass, and three middle voices. Graphical representation of the data will be seen later. I listened to Bowers & Wilkins MM-1 speakers without headphones.


Measuring performance details (Gramophone)

My task was to measure a performance from a sound file. The details needed were: tempo and tuning, start and end times of each sounding note, loudness of those notes, sustaining pedal change times (depressing and raising) and depth (degree of partial pedal), and soft pedal times on and off (it could be assumed that the sostenuto pedal was not used here). Some features of the sound quality of the instrument used in the recording could also be relevant.

In recent years, efforts have been made to develop an algorithm for automatic extraction of details from a sound file (Mauch, 2010; Melodyne). However, the degree of detail required in the present task is beyond the present capability of algorithms. Even in the future, full success with algorithms seems unlikely, especially when the source is of somewhat poor quality, as it often is with historical recordings. Further, the joint effect of the multitude of interacting factors depends on the peculiar properties of human hearing, so the extraction of numerical details from a sound file would not provide all that is needed for reconstitution. The only resource capable of making the required measurements is therefore the human ear. Although measuring a performance simply by listening to it might at first seem to lack a scientific basis, its results can be checked by playback for comparison with the source, and it turns out that the method of successive approximation is in general capable of convergence to a very satisfactory result.

The aim in measuring the performance is not to deduce every action that the performer made. A particular piano with its particular regulation had been used to make a recording by particular recording means. Today, we have almost no knowledge of that piano and its

[9] Alternatives, not used here, were a midi-controlled physical piano or a software implementation with sampled sounds.


regulation, and little knowledge of those recording means and their regulation. Thus the questions exactly what the performer did on his piano and what the recording engineers did with their equipment cannot be answered with any confidence. [10] The question instead is what can be done on a presently available piano, in this case a virtual one, and with presently available technical means, to produce a result which is similar or analogous to the original. Thus not only objectivity but also subjectivity is present and one can hope to achieve, not the largest possible number of automated results of questionable quality, but the best possible result in the single case to hand. Thus science and art cooperate.

Tempo and tuning

All timing measurements require a prior decision on the source playback tempo. Although tempo and tuning level are proportional for gramophone records, that relation cannot be applied here because neither tempo nor tuning level is known with certainty. There was no standard tuning level in Pachmann's time; it varied between and within countries, and the tuning of a particular piano long ago is generally not known, for the opportunity was rarely if ever taken to record at the same time a test signal of known frequency. In this situation I decided to prepare all realizations at today's standard A = 440 Hz; although lower tuning was likely at the time of the recordings, a lower tuning may sound a little dull to today's listeners, an effect which I thought was better avoided (the tuning level can of course easily be adjusted subsequently if desired).

It is also necessary to decide, separately, on the tempo of the realizations. The tempo of the operations taking place in recording and playback is not known, and could vary even within a single take. Different commercial releases of transfers of a gramophone record have different tempos, confirming the uncertainty in this area. For the gramophone record I decided to use the tempo of the Dante CD, which has the advantage of allowing direct comparison of the reconstitution with the source. Fortunately, slight tempo differences do not greatly affect the impression made on listeners.

Note start times

I first measured all the note start times. For this purpose I slowed the sound file from the Dante CD to 20% (for trills 4%) of its original speed while maintaining its pitch, using the commercial computer program Amazing Slow Downer. Listening to the slowed sound played back in the commercial computer program Goldwave, I located the next note played. By starting the playback from successive times just before and just after the note started to sound, I read off the starting time. [11] The loudness level of the playback was maintained throughout, to avoid introducing bias.

That apparently simple procedure conceals several problems. The slowing-down process cannot be perfect and will produce artefacts which, however, are likely to have fairly similar behaviour for the various notes in the piece. The starts of notes that are low or

[10] It should also be borne in mind that Pachmann's piano-playing technique is known to have been very individual, so one should not assume that the technical details that may arise in one's own playing can be transferred to his playing.

[11] The resource of filtering the original sound to highlight a target note was available, but it would have complicated the procedure, and I did not use it.


soft are relatively hard to determine. When several notes start nearly simultaneously it is more difficult to focus attention on the start of the one currently being measured. The starts of each note of an octave or other closely related interval, or of a small interval, are sometimes hard to distinguish because of masking. However, the harder a note is to hear, the less accuracy is needed for successful reconstitution. The task of focusing the attention is in general not a trivial one, and the degree of success in focusing influences the reliability of the measurement. [12] The clerical task was also considerable, with 1,207 notes present but, again, any significant errors would be heard on playback.

Despite these and other difficulties the degree of success, which was judged by playing back the resulting midi file, was found quite satisfactory. The comparison of the reconstituted sound with the source sound as subjectively perceived is the touchstone, not the objective accuracy of the measurements; but from multiple independent timings I had the impression that the objective accuracy was usually within 10 or 20 ms, depending on the sound context. I recorded each timing with a resolution of 1 ms for, even though no one measurement could be so precise, a pattern of successive timings might be distinguishable on those terms. The timing of note starts, with repeated checking, took about 10 minutes per note.

After completing the measurement of the note starts for the whole piece I turned to the remaining three categories. All three had to be measured jointly, for they interact with one another. I made a first estimate of the pedal use, possibly to be revised later, then the note end times (thus the articulation), and finally the loudness, completing that triple procedure for one bar before moving on to the next bar. This work took about three full days per bar. The three categories will now be discussed in order.

Sustaining pedal

The fine control of the sustaining pedal, often abbreviated to just the pedal, is vital to artistic performance. [13] It is successively depressed and raised, with a timing and depth to be estimated (when the source is a sound file, as here, the term 'estimation' seems preferable to 'measurement'). Its timing has an effect in relation to the timing and loudness of the surrounding notes played, while the proportion of the full depth at which it is held, called partial pedalling, determines the size of its effect via the vertical positioning of the dampers in relation to the strings. It would be difficult if not impossible to deduce the pedalling details by statistical analysis of the wave form, and it is not always easy to estimate them by close listening. Any help that could be provided from scores is likely to be inadequate: an experienced pianist often feels little need to adopt editorial pedal indications, for they are imprecisely notated, are sometimes provided mainly for the guidance of amateurs, and depend on the particular instrument used and the acoustic conditions. The pedal indications in the Pachmann editions use the usual imprecise symbols. The present estimates were

[12] The task of focusing on the start of one note in the presence of others is related to the well-known 'cocktail party' effect in the psychology of hearing.

[13] Curiously, pedal is not mentioned in the section on "Performance Data from Recordings" in Clarke (2004, pp. 88-89). Further, Clarke wrote that midi data could not be obtained from historical recordings (p. 87). I make those observations not as criticisms, but to indicate the new departure of my present research. (In the same chapter the word "sostenuto" is used in error for "sustaining" [pp. 80, 81]).


chosen to produce an effect on the given virtual instrument corresponding as closely as possible to the effect heard on the gramophone record with its physical instrument. The extent to which those estimates correspond to Pachmann's physical foot motions cannot be known but is in any case immaterial, just so long as the effect matches. It was nevertheless convenient to estimate in advance the quickest pedal change (raising and depressing) feasible for a human player. That time depends on the depth of the pedal before and after the change, as well as on the regulation of the piano, but I took as a generally applicable value 200 ms, consistent with Repp (1997, Figure 3, p. 168). [14]

My method was to listen to the music at various playback speeds while maintaining pitch, estimating the pedalling by trial and error, comparison between reconstitution and source being readily to hand via playback with the Matlab program. In this work it is noticed that a small numerical change can produce a large audible effect. The pedalling in the original seemed sometimes not quite clean; [15] in such cases I aimed to reproduce the original effect, and did not presume to 'improve' it. In other cases no pedal change effect could be heard, yet it may have taken place and been covered by the surface noise or 'historical haze'; [16] in such cases, whether a change was included in the reconstitution or not would make little or no audible difference, but I generally included a small change. [17]

For partial pedal, the midi specification provides 127 levels. Devices producing sound output from midi files often implement only two or three of those levels, but the Pianoteq software used here implements all 127 levels. The level of partial pedalling heard cannot, of course, be estimated with such accuracy, but within a suitably reduced resolution I was able to recreate reasonably well the effects heard in the source file. Pianoteq takes into account the free play of the physical pedal before the dampers begin to lift (using a default value of 20% of the total travel) and after they reach their full height (25%). Although this was not important when specifying the pedalling through software rather than with a physical pedal, I nevertheless adhered to it for convenience; thus only the intermediate 55% of the travel was effective. In the Pachmann editions (and almost all others) partial pedalling, though it was undoubtedly used, is not indicated at all.

Soft pedal

The soft or una corda pedal has a different degree of effect on different pianos, and possibly a different kind of effect, depending on its design and regulation. On a physical grand piano, it is intended to be depressed fully, relocating the action, and thus the hammers, horizontally in relation to the strings. Pianoteq allows the extent of the full effect of the soft pedal to be specified, and it also provides 127 levels of 'partial soft pedal', by which is meant a fraction of the full sound effect, not a fraction of the physical travel of the action. 'Partial

[14] Repp's experimental situation was considerably different from the present recording one (Repp used a different kind of piano with unknown regulation, and partial pedalling is mentioned only non-specifically on p. 165), but the value still seemed suitable for my purpose. Incidentally, Pachmann had short legs and sometimes used a special pedalling device on that account.

[15] For instance, at the beginning of bars 6b, 18a, 20a, 33a, 35b, 42a, where a and b refer to the first and second half of the bar.

[16] For instance, in the middle of bar 1 and in similar situations.

[17] I occasionally returned to revise the gramophone pedalling after considering the Welte roll pedalling (to be discussed), but the Duo-Art roll pedalling was considered inadequate for this purpose.


soft pedal' is, then, a concept realized only in software and not on physical pianos. In the present gramophone recording soft pedal was possibly used just twice, for bars 22-30 and 47-57, as indicated explicitly in the Klindworth edition of the score with the words 'una corda'. However, it is not quite certain that Pachmann would have used the soft pedal even there. In the Pachmann editions (which do not include the present Nocturne) soft pedal is never indicated, and Pachmann is known to have disliked the soft pedal sound quality in general. Whether it was used in the present recording was not easy to determine by listening, and in any case it did not have a large effect. I used Pianoteq's default (20%) for the full effect, and applied it with midi value 127; thus I did not use 'partial soft pedal' at all.

Note end times

Three kinds of finger articulation are distinguished, whether the note progression is in the same voice or not: staccato, legato and legatissimo, that is, detached, abutting and overlapping. Legato and legatissimo can be achieved with the fingers alone, or with the use of the sustaining pedal, or by both means. When the sustaining pedal is used, as it is almost throughout the present piece, the key release time cannot be determined accurately from a sound recording. [18] Whether the sustaining pedal is used or not, the end time of a note's sounding cannot always be determined accurately either, because a sound at a given pitch may be produced not by the corresponding key but as an upper partial of a lower sounding note. One must nevertheless attempt to match the effect in the source as closely as possible, taking into account markings in the score, the span of the hands, and likely performing details. Thus from bar 2 the bass voice is marked in the Klindworth edition sempre legatissimo, and so was likely to be to some degree both finger-legatissimo and pedal-legatissimo when physically possible, but finger-staccato and pedal-legatissimo otherwise. By trial and error I settled on a general policy of 20 ms or 30 ms legatissimo for the bass voice (this often turned out to be challenged later by the evidence of the rolls, to be discussed). Legatissimo of either kind was less likely within the top voice, because that voice represents the singing of a melody and legatissimo singing is impossible. For any voice, finger-staccato with pedal-legatissimo was sometimes dictated by the configuration of the notes in relation to the hand span, while keeping in mind the possibility, of which Pachmann was very ready to avail himself, of continuing the voice of one hand with the other hand, whether or not this was indicated in the score. I often applied an exact finger-legato to the top voice's melody, but a slight finger-staccato and pedal-staccato representing 'breathing' was sometimes used. Strongly expressive nuances in the vicinity of the downbeat (at the bar line) often implied a small break in the sound. Indeed, a brief silence before a new attack can increase the impression of strength in the attack as well as the clarity of the newly struck note, and this often took place in the bass or treble voice. Further evidence on articulation was provided by rolls, to be discussed later. In any case, the audible distinction involved in many of my decisions, in this and in other categories, was slight or negligible.

[18] For instance in bar 56 where the score has a right-hand chord followed by a rest, with pedal throughout.



(i) Resolution: For measuring loudness, the required resolution must first be determined. The 127 levels provided by standard midi were found to be inadequate, for they fail when the loudness has to be shaped sensitively in soft music: proceeding from a note with midi level 5, for instance, allows moving only to the neighbouring levels 6 or 4, a change of 20% or 25% respectively, which does not generally produce sounds sufficiently finely graduated for the desired artistic effect. Fortunately, the midi specifications have recently been extended with a 'high resolution velocity prefix' providing 16,256 levels (numbered from 128 to 16,383), quite enough for the musical purpose. Loudness is, of course, not perceived as a midi level, and I found it most convenient to use a scale of 0.0 to 100.0, allowing one decimal place, to represent the measurement made by my ear, then converting the heard value to a midi level. [19] The conversion from heard loudness to midi levels is not a trivial matter; by listening, I estimated several control points on the conversion function and fitted a cubic function to those points. The result, while not ideal, was adequate for the purpose, especially as the musical output was again to be checked against the source by listening. The maximum loudness in the piece can be estimated in advance, and is relevant in making sure that the loudness data is well spread out over the available numerical range. My task, then, was to listen to each note and assign it a loudness value on the above-mentioned scale. An attempt to establish reference levels would fail because those levels would depend on pitch, increasing the required number of reference levels beyond what would be practicable. With experience built up by trial and error, I was soon able to make likely first approximations, then to be refined. All checking of the reconstituted version against the original faces the difficulty that the two versions necessarily have considerably different sound qualities, and a consistent attitude needs to be taken to this problem.

(ii) Comparisons: The absolute loudness level of the original playing is not known, and we can only turn our playback volume knob to suit ourselves. The relative loudness of each note is therefore estimated by means of comparisons between the loudness of various notes. So many comparisons could be made that it is impossible to pay attention to all of them over a long stretch of multi-voiced music. [20] One therefore chooses those comparisons that seem most important; in the present piece comparisons may be made between successive low notes at the bar and half-bar, between the several notes of a melodic fragment, and so on. A judge of my estimates might well hit upon an unfavourable comparison, leading to an adjustment for an improved result, but it is to be borne in mind that any one adjustment may have far-reaching effects on other loudness relationships. The many interactions with other parameters also play a part; thus a note starting after a silence (perhaps from staccato) will sound louder than if it had started without a preliminary silence. The presence and degree of sustaining pedal also influences the perceived loudness, as does the soft pedal (despite the conventional wisdom that the latter affects only sound

[19] I thus worked with a resolution of 1,000 loudness values mapped to 16,256 velocity levels. By comparison, Zenph used the Disklavier Pro piano in XP mode, which has 1,024 velocity levels. As with timings, it was considered best to record one more decimal place than appropriate to any one measurement.

[20] For instance, the present piece with its 1,207 notes allows nearly 700,000 paired comparisons and 300,000,000 three-way ones.


colour). Comparisons will most often be made between notes close together in time, but one also needs to check that a drift or other change has not arisen in more distant comparisons. For distant comparisons one has to compare suitable excerpts from both the reconstituted file and the original file and then make any required adjustments; this process is time-consuming but essential.

[21] Comparisons of loudness between voices may be difficult to make when the pitch range is large; the balance between voices must of course nevertheless be checked. Adjustments for balance usually have to be made manually, that is, a revised loudness is typed in separately for each affected note, and this too is time- consuming. No matter what balance is achieved, it is enough that it is consistent throughout the piece, for then listeners to the final result may and should adjust their equalizer according to the properties of their own playback system.

(iii) Chords: Comparison of the loudness values of the component notes of a near- simultaneously-played chord is an especially difficult problem, which could well be the subject of future research. In estimation based on such comparisons the psychological category of 'attention' comes into play, as it does also with many other estimation tasks in this work. Repeated listening to any musical fragment may result in different estimates depending on where the attention happens to be focused, as well as the temporal order of the focusing. Attention may vary in the frequency dimension or in the time dimension. Thus no two listeners, even if their hearing is in other respects similar, are likely to hear in the same way, nor is any one listener on different occasions. After close listening, focusing on each note of an excerpt in turn, it may therefore be best to replay the whole excerpt for a perspective over a wider time-span and a less narrowly focused frequency range. No other practicable remedy seems to be available for this difficulty. The loudness of the two notes of an octave is perhaps the most challenging case. The loudness of the lower note might hardly be capable of good estimation by attempting to pay attention just to that lower note, for it will be partly masked by the upper note which, in turn, will absorb an upper partial of the lower one. In such cases, the effect of the whole can be approximated by trial-and-error adjustment of the two loudness values, providing a serviceable indirect estimate.

Compensation for loudness attenuation

The only extraneous improvement to the recording that I considered desirable was to avoid the well-known loss of loudness in high and low notes brought about by the recording process.

[22] Not only loudness but also apparent articulation is affected, for such notes may sound shorter in the recording than in the original performance. I first attempted to reconstitute the sound as heard in the source and then, as a separate step, to adjust for the mentioned loss of loudness. I have attempted that adjustment only in an informal way, for the available documentation of the properties of the recording process is not adequate for a

[21] One might consider comparing the loudness outlines of the original and reconstituted versions by plotting their smoothed wave files. However, the different sound qualities can have a strong influence, so this method is likely to be inadequate.

[22] This is called a frequency response correction. I did not find a need to compensate for dynamic range compression, which might more likely be needed for recordings made by the acoustic rather than the electrical process, that is, for earlier recordings than the present one.


formal approach. I used a piece-wise linear function (Figure 2). In the present composition this compensation could have a noticeable effect only on some of the lowest bass notes and on occasional high notes as in bars 37-38. My procedure assumes that such an adjustment had not already been carried out in the record mastering, in this case for the Dante CD; but in any case the final result must reflect the judgement of an individual. [23] After such loudness boosting, the notes then also sound longer, and no separate length adjustment was made. The resulting reconstitution is heard in Audio 2 ℗

Audio 2:         (Or click here for an alternative playing method, then the browser's back-arrow.)

Figure 2. Factor for loudness compensation for each pitch.

Commentary on the playing (Gramophone)

The large amount of data now available from the chosen performance allows several kinds of commentary or analysis. A "nuanced score" (Nettheim, 2001a) is an instructive resource, intended for future preparation. Analysis of loudness patterns would, on the other hand, be of limited use because the estimated values determine loudness not alone but in conjunction with the articulation of the previous note, as well as with pedal and the musical context. The level of adherence to the expression markings in the score could be evaluated, but will also not be pursued formally here; informally it was found to be high. Some other details are now spelled out in three categories: divergence from the score, trills, and miscellaneous observations; those are followed by a more novel feature.

[23] This process could instead have been carried out in Pianoteq via its 'velocity map', but the way I chose retains control even if another output form is later used.


(a) Occasional slight divergences were found between the notes of the performance and of the Klindworth score. (i) Several notes did not sound for accidental reasons. [24] Whether the lack of sound resulted from depressing the key too gently or in the recording process is uncertain; in any case, not pedantry but only documentation is intended here. (ii) Conversely, one note did sound for accidental reasons, and I implemented that slight blurring. [25] A 'phantom' note may also occasionally be heard. [26] (iii) The order of the notes in the score was sometimes changed. [27] (iv) Deliberate changes from the score notes seldom occurred. [28]

(b) The playing of the rapid trills in bars 36-37 reveals, in brief, pairs of notes played so closely together within each pair that the technique used may have been to hold the two fingers with one raised slightly higher than the other, playing with a single downward motion of the arm or hand for each pair of notes. [29]

(c) A few miscellaneous observations may be made. In bar 5 a curious slight delay occurs before the last three notes, presumably while changing fingers on the last B2. In bars 32 and 33 Pachmann played the sextuplet as three pairs, not two triplets as indicated in the Klindworth score. The last B3 in bar 56 was played suddenly much more loudly than its surroundings, which may at first seem puzzling, but Pachmann was following the marking in Klindworth's score indicating, not only with a separate stem but also with a slur, a prominent two-note progression from this note to the final E4.

The main analytical feature I wish to present here is a novel one, concerning the treatment of the melody compared with the remainder of the musical texture. I have previously written on that topic (Nettheim, 2001a, in press) in respect of just a few bars of this music, but a much fuller demonstration of it is available here. I accordingly separated the melody (Audio 3 ℗) and the bass (Audio 4 ℗) from the remainder of the music (allowing the pedal to remain as in the whole of the music). This shows highly regular playing of the melody with nothing resembling eccentricity, just as a singer might deliver it. On the other hand Pachmann played the bass, I suggest, not primarily with a view to its own internal

Audio 3:         (Or click here for an alternative playing method, then the browser's back-arrow.)

Audio 4:         (Or click here for an alternative playing method, then the browser's back-arrow.)

[24] Bar 21 2nd C sharp 3, bar 30 1st A3, bar 36 several notes in the trill, bar 41 last G4, bar 45 last B1, bar 49 last E3.

[25] Bar 12 the first note D sharp 2, blurred by its neighbour E2.

[26] This may result from a low note whose upper partials are subsequently reinforced by the playing of other notes, as for instance bar 46 at about the second B2, where the low E1 may be heard as if played again, in both source and reconstitution. Thus the ear may occasionally need to be aided by the score, by musical logic, or by both.

[27] Bar 11 last B3, bar 12 1st B3, bar 16 last A, bar 17 left hand F sharp 4, last E3 and G4 (this is the most curious case here, not matched in the rolls), bar 31 last B3, bar 35 G3, C sharp 4.

[28] Bar 8 a second soft C sharp 5 added (discussed later); bar 16 G4 and bar 17 A4 omitted (also omitted from the Welte roll but not from the Duo-Art roll, suggesting though not proving editorial interference in the latter case); bar 54 G sharp 4 and B4 repeated (as if to make those voices clear, but not appearing in the rolls).

[29] This technique may be the one mentioned in Fleisher (2010, p. 48), in connection with the teaching of Karl Ulrich Schnabel: "...playing both notes together and then starting to lift the alternate finger so that it sounded a fraction later, so that you start out by playing an interval and gradually separate it out into two distinct, alternating tones."


coherence, but so as to enhance the melody by remaining subservient to it. The accompaniment steps aside where needed to allow a melodic tone to shine forth without having other notes compete for the listener's attention. [30] This challenges the popular dictum that the left hand should keep time while the right hand plays freely, for the opposite occurs here. [31] Yet Pachmann did not take the approach observed here in all cases of lyrical piano playing, as my preliminary study of the middle section of the Chopin Waltz op.64/2 shows (work in progress).

Conclusions (Gramophone)

For general listening, a transfer of the original 78 rpm recording is no doubt more authentic than the present reconstitution (except for the loudness attenuation in the former), but the different sound qualities might influence a listener's preference: the 'historical haze' is here exchanged for clear modernity. For a listener seeking instruction, the present approach, if it has been carried out well enough, offers definite advantages. In particular, the playing of separate voices of the piece by themselves seems very instructive, and such a resource has perhaps never before been made available from a gramophone recording as source.


Piano rolls have generally received a negative response from connoisseurs of recorded music, [32] so it is worthwhile trying to improve the results obtained from rolls. The shortcomings of rolls and possible remedies for those shortcomings will be discussed in the following sections on measurement.

The Welte factory at Freiburg, together with the secrets of its technology, was unfortunately destroyed during World War II; instead of knowledge of its method, what remains is all or largely speculation, and researchers therefore have to work with incomplete and uncertain information. Further, the likely work of the original editor of each roll introduces a pervasive element of doubt.

Various methods have been used for digitizing the holes punched in a roll. I used the custom mechanical roll reader built by Peter Phillips of Sydney, Australia. [33] His reader produces a midi file which may then be passed through a third-party software 'emulator' to realise the intended effects of the 'expression tracks' for loudness and pedal (the emulator takes account of the special mode of operation of rolls, including built-in delays needed for the pneumatic system).

[30] Compare Adelina Patti's remark to Pachmann, reported in Blickstein (1969, p. 59): "I sing with my throat, but you sing with your hands, which is even more precious."

[31] "According to a current story, Chopin used to say to his pupils: 'Play freely with the right hand, but [let] the left one act as your conductor and keep time.' We do not know whether the story should be afforded the benefit of the doubt" (Paderewski, 2001).

[32] A partial exception is Newman (1920), but Newman's defence was not made in terms of playing at the highest level: "If it be urged that the piano-player [of rolls] cannot produce all the exquisite artistic effects of a Pachmann, we cannot but agree" (pp. 30-31). Other apparent exceptions are seen in the advertisements in which recording artists promoted their rolls, but these may well have been influenced by a contractual or financial interest.

[33] Phillips is currently writing up the details of his reader as part of a thesis at the University of Sydney.


Measuring performance details (Welte)

We turn now to the Welte roll (Chopin, 1827/1906 or 1925). A commercial transfer of the original is provided as Audio 5. The same performance details need to be measured for rolls as for the gramophone record discussed earlier, but the methods of measurement are very different.

Audio 5:         (Or click here for an alternative playing method, then the browser's back-arrow.)

Tempo and tuning

Welte rolls indicate a tempo in feet of paper per ten minutes. At the present roll's tempo of 75 the duration, measured from the start of the first note to the start of the last note, was 242.5 s, compared with the Dante CD's 236.1 s, the difference of 2.7% between performances being quite believable. I was not aware of a problem related to acceleration of the roll while being played that sometimes occurs (Stahnke, 1999, pp. 11-12). I set the tuning at A = 440 Hz as I had done for the gramophone reconstitution.

Note start times and end times

The note start times punched on rolls may be generally reliable, as is consistent with the heard results to follow, although the notes themselves might occasionally have been edited to correct a note assumed to have been wrongly performed. Note end times may also be reliable, but in some cases they seem, on the basis of their being unplayable by human hands, to have been edited to produce a longer note, presumably to assist with a legato effect not always easily obtained with the roll's pedal mechanism. Note start and end times may be less reliable with very short notes, including those in trills, because of mechanical limitations. When two notes start within a few ms of one another, the note order as read may be reversed, which however did not correspond to an audible effect here (on the other hand, larger departures from simultaneity can be important, as in Goebl et al., 2008). Preliminary testing of the roll reader by means of multiple readings had suggested a consistency generally within 20 ms, depending on the musical context. This mechanical method of measurement saves many months of work per roll compared with auditory measurement from the sound of a transfer or by visual measurement of the positions of the holes (simple visual measurement would not alone be adequate because of the subtlety of roll technology).

Sustaining pedal

An important shortcoming of rolls is their limited specification of the sustaining pedal. The specification is encoded near the edges of the roll but the reliability of the recording mechanism is not certain, and the encoding may have been altered by an editor. The pedal start and stop times are thus questionable, and partial pedal is not specified at all, only fully on or fully off. [34]

As a first step, I used the sustaining pedalling indicated on the roll. This gave generally fairly acceptable results here, despite the above reservations. The most common pedal change duration (for raising and depressing the pedal) was 500 - 600 ms, greater than the

[34] All comments on rolls made in this paper refer only to those manufactured by Welte and Duo-Art, a consideration of the rolls of other manufacturers being beyond the present scope.


200 ms I had used for the gramophone reconstitution, but I subsequently reduced many of the roll's pedal change durations to 200 ms. The sustaining pedal change durations that Pachmann used cannot be known; [35] but that is not a serious limitation, for they would in any case not necessarily be appropriate on the virtual instrument now used. I determined the degree of partial pedal, a vital contribution to the artistic result, by trial and error. A systematic study of partial pedalling would be a difficult task requiring a large number of sound examples, because of the innumerable contexts in which it may be used (Schnabel, 1950, pp. 14 & ff).

Soft pedal

The soft pedal occurred on the Welte roll in many short bursts that would be unnatural in real playing. They may have been be supplied by an editor to allow some air to escape from the reproducing mechanism in certain situations. I changed them to the same two instances I had used for the gramophone record.


The limited scope for specifying loudness on rolls is one of their greatest shortcomings. [36] Holes punched at the sides of the roll indicate loudness, but whether they were recorded at the time of the performance or added (or modified) later by an editor is generally uncertain. Further, there is a limit on the softness achievable with a roll, because the playback is pneumatic and a certain minimum air flow is needed to produce a sound reliably; as a result, there is also a limit on the fine shading of soft notes on a roll. The loudness achievable may also be limited, for mechanical reasons. Very importantly, the loudness of individual notes within a chord whose notes are sounding simultaneously cannot generally be specified on a roll. Various methods for specifying loudness have been used by different roll-producing companies, but if two or more notes are recorded closer together than a certain minimum separation time, then in a pneumatic system they are all played with the same puff of air, so their individual loudness levels cannot be controlled.

To remedy this, I took as a first approximation the loudness pattern I had measured from the gramophone recording. [37] This step was reasonable because the two performances were generally fairly similar in other respects. [38] In the few cases when notes present on the roll had not sounded on the gramophone recording, I estimated the loudness freshly, bearing in mind the musical context. When significant differences were heard between the

[35] The 1923 silent film clip (unidentified repertoire), on which Pachmann's pedalling can at some places be seen, shows a slow pedal change early, perhaps 500 - 600 ms, and later apparently much quicker partial changes.

[36] "The pianist Artur Schnabel was approached by one of the companies that manufactured piano rolls early in this [20th] century and invited him to record for them; 'Our system,' they proudly told him, 'can reproduce 16 different degrees of touch.' 'What a pity!' Schnabel replied. 'You see, I have seventeen' " (Barker, 1999).

[37] To seek improved loudness patterns for rolls of works which the same performer has not also recorded for the gramophone, a different approach would be needed, possibly involving a gramophone recording of similar style by another performer or else freshly supplied values. Such approaches would clearly be less desirable than the present one.

[38] When Pachmann recorded other pieces more than once, the corresponding performances were again generally of similar character.


Welte bar 1

Figure 3. Welte roll reconstitution, bar 1 (the 12-note bar was played twice on the roll). Note-names C are coloured yellow. Notice the variety of loudness specifications I applied, indicated by numbers above the start of each note (0.0-100.0). Sustain pedal is shown under the notes and soft pedal (if any) in the green area below that.


musical gestures of the gramophone recording and those of the roll, as indicated by the timing, I modified the loudness values accordingly. The fresh estimation or modification of loudness requires the judgment of the researcher, which fortunately had to be applied only on a limited number of occasions. On such occasions I adopted a conservative approach by not setting out to conjecture Pachmann-like nuances but merely to follow the implications of the observed timing patterns. A few examples of such modifications follow.

Example A: In the special case of the double-length bar 1 (Figure 3) I made a number of loudness modifications in an effort to achieve a more artistic result than the roll provided.

Example B: In bar 4 (Figure 4) the pattern of note lengths for the descending thirds in the last half of the bar is different from the pattern in the gramophone recording: they are more clearly shaped in dissonance-resolution pairs in the roll performance, so I softened the resolution thirds (B4 & G4, G4 & E4) to loudness values 15.0 & 12.0, 8.0 & 5.2, whereas the gramophone performance treated them more nearly equally with their preceding dissonances, with loudness values 19.0 & 16.0, 9.0 & 6.2.

Bar 4 for each source

Figure 4. Reconstitution, bar 4, two voices. Top: gramophone; middle: Welte; bottom: Duo-Art. Notice the different relative durations of the successive intervals and the correspondingly different relative loudness values.

Example C: An interval might be played with slight arpeggiation in one version but not in another version. The anticipating note might be intended as leaning into the following main note, in which case the anticipating note might be softened, or as being highlighted, in which case it might instead be played louder. An example where softening is likely occurs at the beginning of bar 16 (Figure 5), where E4 anticipated A4 by 65.806 - 65.775 = 31 ms in the Welte version, so I changed its loudness from the gramophone value 20.0 to the Welte value 10.0 (in the gramophone, the anticipation had been much larger, 59.772 - 59.554 = 218 ms, changing its category and therefore its manner of delivery).

Example D: In bar 56 the B3 was left with loudness unchanged from the gramophone recording, again reflecting the nuance indicated by Klindworth.


Bar 16 for each source

Figure 5. Reconstitution, bar 16. Top: gramophone; Bottom: Welte roll. Notice the anticipation of the first A4 by E4: long in gramophone, short in Welte where the E4 is therefore softer (10.0 instead of 20.0). The voices are distinguished by their colour.

The full verbal description of loudness modifications, let alone pedalling modifications, would be too extensively detailed for the present article, but the above examples may suffice. The way in which a change in a timing pattern was found to imply a certain change in the corresponding appropriate loudness pattern was a most interesting feature, reflecting the indissoluble linking between timing and loudness and calling for further research towards a more systematic exposition. It was fortunate that only a fairly small number of such adjustments, on average about one per bar (apart from bar 1), seemed needed for this roll.


Welte bars 40-42
Figure 6. Welte roll reconstitution, bars 40-42. Notice the legatissimo spanning wide-spread pitches in the lower part.


Welte bar 31
Figure 7. Welte roll reconstitution, bar 31. Notice the long-held G5.


Compensation for loudness attenuation
The loudness estimated directly from the gramophone recording was not used here, but instead the values obtained after applying loudness compensation; no loudness compensation was therefore applied to the roll, for it did not give rise to attenuated dynamics. The reconstitution is provided as Audio 6 ℗.

Audio 6:         (Or click here for an alternative playing method, then the browser's back-arrow.)

Commentary on the playing (Welte)

A study of articulation on rolls could throw light on the fingering used by the pianist, but this will not be pursued here; it will be assisted when studying pieces for which there is a Pachmann edition including his fingering. Some other details follow:

(a) (i) Any notes that did not sound in the performance for accidental reasons are likely to have been supplied by the roll's editor. It was tempting to 'humanize' the rendition by seeking to reverse such presumed editing but, in the absence of a log of the editing, I did not make that attempt.  (ii) Any blurred notes are also likely to have been removed by the roll's editor.  (iii) The order of the notes as performed on the roll is compared not directly with the score but with those in the gramophone recording; several differences were found. [39]   (iv) The most obvious deliberate change from the score is the repetition of bar 1. The measurements of the two playings of that bar are not identical, so the repetition is almost certainly not the result of an editorial action, and the explanation of it is not known; I have retained the repetition in the reconstitution. [40] Pachmann also varied from the Klindworth score in bar 29, where an extra A sharp 4 was played in the middle of the bar (not in the gramophone or Duo-Art roll), and in bar 53 where an extra D sharp 5 was played in the middle of the bar (again not in the gramophone or Duo-Art roll), in both cases no doubt played softly but gaining extra sounding length for the long note. Combining those two cases with bar 8 of the gramophone and Duo-Art roll, thus altogether bars 8, 29 and 53, we see that these three are the only whole-notes in the score; Pachmann has given them extra sounding-length beyond the piano's naturally decaying tone, an instructive conception which reminds us once again that a literal-minded approach to score-reading is (or was) by no means required. Finally, in bar 57 the score's E3 was not played.  (v) The articulation, that is, the ends of notes, could scarcely be studied from the gramophone recording, but can be studied from the roll. In the double-length bar 1 (Figure 3) the articulation three times implies that Pachmann was comfortable spanning all three notes of the E2-B2-G3 chord. Considering his small hand, that seems unlikely even with narrower than standard keys, and already suggests, though it does not confirm, that the end times had been adjusted editorially. If both hands were used in that bar, the progression to the next bar would be compromised. Similar figures in some other bars also have such legatissimo; in some cases the right hand might have been used for the top note of the figure, but in bars 40 and 42a some spans seem difficult or impossible (Figure 6). In bar 31 (Figure 7) the long- held G5 seems to inconvenience the playing of the following short notes. Taken together, such evidence suggests editorial intervention.

[39] Bar 3 F sharp 2, bar 10 G4, and bar 17 the last two notes G4, E4 which, unlike the gramophone recording, follow the score order (conceivably after editing).

[40] Pachmann played a similar repetition of the opening bars in his two Welte roll recordings of Henselt's La Gondola op.13/2, called 'Improvisation "En forme de Gondola"'


(b) The trills in bars 36-37, on slow playback, give less impression of a pairing of the notes in separate gestures than did the gramophone record; but the reliability of the roll data in such cases is uncertain.

(c) Miscellaneous observations: in bar 5 the delay in the gramophone recording is not present here; in bar 56 the nuance on the last B3 in the gramophone recording is probably present again here.

As with the gramophone record, so also on the Welte roll the melody alone (Audio 7 ℗) is played in a straightforward manner, while the bass (Audio 8 ℗) steps aside whenever required to allow the melody to shine through.

Audio 7:         (Or click here for an alternative playing method, then the browser's back-arrow.)

Audio 8:         (Or click here for an alternative playing method, then the browser's back-arrow.)

Conclusions (Welte)

The reconstitution of the Welte roll seems a clear improvement over the roll as originally published. It seems fair to say that, until this reconstitution was carried out, the performance had never been heard in a manner approaching the way the artist had played it when making the roll. I made some decisions for the reconstitution myself but sought to minimize those and to take a conservative attitude with them; it should also be borne in mind that the source roll had itself been subjected to editing to an unknown extent. Separating melody and bass from the whole musical texture was just as instructive here as in the gramophone recording. To review the aesthetic quality of the source performance as here reconstituted is not an essential function of this article but rather something for the reader to carry out after listening; on that understanding, I offer my opinion that it is in some ways even more beautiful than the gramophone performance.


We turn finally to the Duo-Art roll (Chopin, 1827/(1923 or earlier). A commercial transfer of the original is provided as Audio 9. [41] As with Welte rolls, the Duo-Art recording and production procedures are not known in full detail.

Audio 9:         (Or click here for an alternative playing method, then the browser's back-arrow.)

Measuring performance details (Duo-Art)

Tempo and tuning
The Duo-Art roll reproduced as a much slower performance: 271.9 sec, a difference of 15.2% from the Dante gramophone duration (236.1 s). [42] While not wishing to negate what could conceivably have been a very different approach by Pachmann, I nevertheless found some of the slow effects to be quite hard to believe. I therefore assumed that an error had been made in the labelling of the roll, and I accordingly changed its 95 feet per ten minutes to 105 feet per ten minutes, bringing the resulting tempo in line with the other sources, and the duration to 246.0 sec, thus within 4.2% of the Dante duration. I again set the pitch at A = 440 Hz.

[41] The sound quality of this version leaves much to be desired, but it serves here to show the content of the roll as played on a conventional reproducing piano.

[42] Comparisons of Duo-Art rolls of other pieces with performances on gramophone or on Welte showed much smaller tempo departures. Compare the discussion of a similar large discrepancy in Stahnke (1999, pp. 10-11).


Note start and end times
These were again measured with the Phillips roll reader, a different 'tracker bar' being required for the different roll format.

Sustaining pedal
By contrast with the Welte sustain pedalling, the Duo-Art sustain pedalling was found inadequate even as a first approximation. I therefore used the pedalling measured from the gramophone recording as a first approximation, introducing modifications according to local contexts when needed. The degree of partial pedal, an important contributing factor, was again determined by trial and error.

Soft pedal
Example (a). Bar 4 ends with four descending intervals of a third (Figure 4, bottom). In the gramophone recording the two notes of each third were played nearly simultaneously (Figure 4, top), but in the Duo-Art roll the C 5 anticipated its companion by 15.331 - 15.242 = 89 ms. The lower voice was surely not intended to be emphasized and therefore needed softening in the roll; the lower voice of the following three intervals then needed to follow suit by also being softened, even though the notes of the last two intervals are played almost simultaneously. This illustrates the matching of expression at one place to that in another place. This was also inadequately specified on the roll, and I handled it as I had done for the Welte roll.

This was again perhaps the greatest shortcoming of the roll. [43] Again I took the measured gramophone loudness values as a first approximation, then making modifications as needed. Two examples of the modifications follow.

Example (a). Bar 4 ends with four descending intervals of a third (Figure 4, bottom). In the gramophone recording the two notes of each third were played nearly simultaneously (Figure 4, top), but in the Duo-Art roll the C 5 anticipated its companion by 15.331 - 15.242 = 89 ms. The lower voice was surely not intended to be emphasized and therefore needed softening in the roll; the lower voice of the following three intervals then needed to follow suit by also being softened, even though the notes of the last two intervals are played almost simultaneously. This illustrates the matching of expression at one place to that in another place.

Example (b). In bar 6 the low F sharp 2 in the middle of the bar is relatively much longer in the roll than in the gramophone recording (Figure 8) and sounds enfeebled when given the loudness of the gramophone recording. From this, it seems clear that Pachmann emphasized that note in the roll performance and not in the gramophone recording, so I gave greater loudness to the roll's F sharp 2, and therefore also to the earlier G2 which forms a unit with it; further, the two greater loudness values required an extra pedal change. This illustrates once more the linking of the length and loudness of the notes, as well as the pedal, in artistic music-making.

As with the Welte roll, an average of about one modification per bar to the gramophone recording's loudness values was made.

[43] "The great pianist Josef Hofmann ... wrote ...: 'As regards the Duo Art rolls...they did well for a while, for at that period the phonograph records were unable to reproduce the piano tone faithfully. ["phonograph" generally refers to a cylinder recording, "gramophone" to a disc recording.] Rhythmically they were all correct; dynamically they were a fake.'   'Fake' may seem a strong word, but in the (sic) fact the mechanisms for recording the pianist could not capture dynamics and nuances of touch – these were added later, often with the help of the artists, but often not" (Barker, 1999).


Compensation for loudness attenuation
As observed earlier, no loudness compensation is required for rolls. The reconstitution is provided as Audio 10 ℗.

Audio 10:         (Or click here for an alternative playing method, then the browser's back-arrow.)

Bar 6 for each source

Figure 8. Reconstitution, bar 6. Top: gramophone; bottom: Duo-Art roll. Notice that the F sharp 2 in the middle of the bar is relatively short in the gramophone but longer in the roll where it has therefore been made louder (7.0, 12.7 respectively).

Commentary on the playing (Duo-Art)

(a) Some of the notes that had not sounded on the gramophone recording did sound on the Duo-Art roll (compare the commentary on the Welte roll). Two slightly more significant changes were observed: In bar 19 the order of the 3rd and 4th notes in the left hand was unaccountably reversed, and I restored the score order. In bar 35 a curious B4 was added at the beginning of the bar; though it may well be an editorial error, I allowed it to remain.

(b) The trills in bars 36-37 have more notes than in the gramophone and Welte roll, but do not suggest the paired technique mentioned for the gramophone recording.


Separation into melody (Audio 11 ℗) and bass (Audio 12 ℗) again confirms the previous findings on this point.

Audio 11:         (Or click here for an alternative playing method, then the browser's back-arrow.)

Audio 12:         (Or click here for an alternative playing method, then the browser's back-arrow.)

Conclusions (Duo-Art)

Even more than with the Welte roll, the reconstitution of the Duo-Art roll seems a great improvement over the source roll material, and again the performance can now perhaps be heard properly for the first time. Separating the melody and bass for individual playing was equally instructive here. Others will judge the artistic quality of the source performance, which I consider very fine.


The reconstitution of three different recordings of a chosen work – gramophone, Welte roll and Duo-Art roll – has been carried out with what may in each case fairly be claimed to be artistic success. The use of the gramophone loudness and pedalling in place of those of the rolls was a novel and helpful feature (which would however not be available for all recordings). Apart from the possible independent value of the reconstitutions to general listeners, the value to students of performance lies in the facilitation of close analysis, and especially the separation of the melody and bass from the whole of the music. The way is now open for the preparation of a 'nuanced score' (Nettheim, 2001a) derived from the gramophone recording, or even from the rolls. The resources used, including midi representation and the Pianoteq virtual piano, were fully adequate to the task. The strong demands made by the processing method constitute a negative feature; they include the large amount of time required for successive approximations and the concentrated use of the human ear. It remains to be seen whether such demands on researchers will be tolerated in future work, by comparison with faster automated methods of adequate quality that are being sought but that do not currently, and might never, exist.

ACKNOWLEDGEMENTS: I thank Peter Phillips for kindly making available his custom-built piano-roll reader. I also thank Werner Goebl, Peter Phillips and Bruno H. Repp for closely reading and commenting on the manuscript, as well as the two anonymous reviewers.

COPYRIGHT NOTICE: Audio examples 2, 3, 4, 6, 7, 8, 10, 11, 12 are copyright ℗ Nigel Nettheim (2013).


Chopin, F. (1827). Nocturne in E Minor. [Recorded by V. De Pachmann]:

(i) [78 rpm record]. London: HMV 2-05567; HMV DB 1106; Victor 6879-A. (1927)

(ii) [Piano roll]. London or New York: Welte Licensee C-7202. (1906 or 1925)

(iii) [Piano roll]. London: Duo-Art 016 reissued as New York: Duo-Art 6675. (1923 or earlier)

(iv) On Vladimir de Pachmann: complete recordings volume 1 [CD]. France: Dante. (1996)

(v) On The Welte legacy of piano treasures: Vladimir de Pachmann performs in 1906


[LP from Welte roll]. Hollywood, California: Recorded Treasures, Inc. (1963)

(vi) On Vladimir de Pachmann plays Chopin [LP from Duo-Art roll]. Los Angeles: Everest Records. (1966) Available {May 18, 2022} at http://archive.org/details/ChopinWorksForPiano

Slåttebrekk, S. (2010). Edvard Grieg: chasing the butterfly. Recreating Grieg's 1903 recordings and beyond [CD]. Presto Classical PSC1299.

[Unidentified repertoire.] [Recorded by V. De Pachmann, 1923]. [Silent film clip]. London: Duo-Art.

Zenph Studios (2007). Glenn Gould. Bach: The Goldberg Variations. 1955 Performance [CD]. Sony 88697-03350-2.

Zenph Studios (2009). Rachmaninoff plays Rachmaninoff [CD]. RCA 88697-48971-2.


Chopin, F. (1827). Nocturne in E minor, op.72/1. In the following editions (* indicates retrieval October 1, 2012, from http://imslp.org/):

Casella, A. (Ed.) (c.1915). Chopin Nocturnes. Paris: Société Française d'Edition des Grands Classiques Musicaux.

Debussy, C. (Ed.) (1915). Chopin Nocturnes. Paris: Durand.

Joseffy, R. (Ed.) (1915). Complete Works for the Piano, Vol.4: Nocturnes. New York: Schirmer.*

Klindworth, K. (Ed.) (c.1880) Chopin. Nocturnes. Moscow, St. Petersburg: P. Jurgenson.

Klindworth, K. (Ed.) (1880). Chopin. Complete Piano Works, Vol. 2. Berlin: Bote & Bock.

Klindworth, T., & Scharwenka, X. (Eds.) (1882). Oeuvres pour le piano, Vol.2: Nocturnes. London: Augener.*

Kullak, T. (Ed.) (1881). Klavierwerke. Instructive Ausgabe, Vol.V: Nocturnes. Berlin: Schlesinger. Also New York: Schirmer (1881).*

Mikuli, C. (Ed.) (1894). Complete Works for the Piano, Vol.4. New York: Schirmer.*

Paderewski, I.J, Bronarski, L. & Turczynski, J. (Eds.) (1951). Chopin. Complete Works, Vol.7. Warsaw: Instytut Fryderyka Chopina.*

Pugno, R. (Ed.) (1902). Frédéric Chopin. Nocturnes. Vienna: Universal.

Sauer, E. (Ed.) (1920). Fr. Chopin Nocturnes. Mainz: Schott.

Scholtz, H. (Ed.) (c.1879). Sämtliche Pianoforte-Werke, Band I. Leipzig: Peters.*

Chopin, F. (Various dates/1934, 1935, 1937). Chopin (piano works) with the authentic fingering and phrasing of Vladimir de Pachmann. Transcribed and with notes by Marguerite de Pachmann-Labori. London: Augener.


Barker, D. (1999, June). [Untitled] [Review of the CD A window in time: Sergei Rachmaninoff, produced by Telarc 80491]. SoundStage! Retrieved January 31, 2012 {May 18, 2022}, from http://www.soundstage.com/music/reviews/rev135.htm

Blickstein, E. (1969, July). More than a clown. High Fidelity Magazine, 19(7), 59-62.

Clarke, E. (2004). Empirical methods in the study of performance. In E. Clarke and N. Cook (Eds.) Empirical musicology (pp. 77-102). Oxford: Oxford University Press.

Fleisher, L. (2010). My nine lives. New York: Doubleday.


Goebl, W., Dixon, S., De Poli, G., Friberg, A., Bresin, R., & Widmer, G. (2008). Sense in expressive music performance: Data acquisition, computational studies, and models. In P. Polotti & D. Rocchesso (Eds.), Sound to sense, sense to sound: A state of the art in sound and music computing (Chapter 5, pp. 195-242). Berlin: Logos.

GoldWave Inc. (2012). GoldWave (Version 5.67) [Computer software]. Available {May 18, 2022} from http://www.goldwave.com

Grabowski, K. (2009). The works of Frédéric Chopin: Score editions and editorial traditions. (W. Bonkowski, Trans.) Retrieved December 1, 2012 {May 18, 2022}, from http://www.chopin.pl/edycja_1999_2009/recepcja/edytorstwo_muzyczne/wydaw_nut_en.html

Malik, F. (2011). Review of the CD Rachmaninoff plays Rachmaninoff produced by Zenph, 2009. ARSC Journal 42(1), 121-123.

Mauch, M. (2010). Automatic chord transcription from audio using computational models of musical context. [Doctoral dissertation, University of London]. Retrieved December 28, 2012 {but not May 18, 2022}, from http://matthiasmauch.de/_pdf/matthiasmthesis.pdf

Melodyne Editor [Computer software]. Munich, Germany: Celemony. Available {but not May 18, 2022} from http://www.celemony.com/cms/index.php?id=products_editor

Modartt (2006). Pianoteq (Version 4.5.1) [Computer software]. Available {May 18, 2022} from https://www.pianoteq.com/

Nettheim, N. (2000). Bibliography of musical scores edited by Vladimir de Pachmann. Retrieved November 6, 2013 {May 18, 2022}, from http://nettheim.com/pachmann/editions/ {and more recently the scores themselves retrieved May 18, 2022 from http://nettheim.com/pachmann/editions/PachmannEditions.html}

Nettheim, N. (2001a). A musical microscope applied to the piano playing of Vladimir de Pachmann. Retrieved December 20, 2012 {May 18, 2022}, from http://nettheim.com/publications/microscope-pachmann/microscope-pachmann.html

Nettheim, N. (2001b). Vladimir de Pachmann, pianist (1848-1933): A discography / rollography. Retrieved December 30, 2012 {May 18, 2022}, from http://nettheim.com/pachmann/discography

Nettheim, N. (in press {when the present article first appeared}). Vladimir de Pachmann and Chopin. In The Third International Congress, Chopin 1810-2010. Book of the congress. Warsaw: The Frederyk Chopin Institute. {Subsequently published as Vladimir de Pachmann and Chopin. Chopin 1810-2010: Ideas, Interpretations, Influence. The Third International Chopin Congress, Warsaw, 25 February to 1 March 2010., Irena Poniatowska and Zofia Chechlinska (Eds). Warsaw: The Frederyk Chopin Institute, 2017, volume 2: 17-26.}

Newman, E. (1920). The piano-player and its music. London: Grant Richards.

Paderewski, I. J. (2001). Tempo rubato. Polish Music Journal, 4(1). (Reprinted from Success in music and how it is won, Chapter 28, by H. T. Finck, 1909/1927, New York, NY: Charles Scribner's Sons.)

Repp, B. H. (1997). The effect of tempo on pedal timing in piano performance. Psychological Research 60(3), 164-72.

Roni Software (2012). Amazing Slow Downer (Version 3.4.1) [Computer software]. Available {May 18, 2022} from http://www.ronimusic.com/

Schnabel, K. U. (1950). Modern technique of the pedal. New York: Mills Music.

Stahnke, W. (1999). [Booklet enclosed with the CD A window in time: Sergei Rachmaninoff]. Telarc 80491.

Zenph Studios (2012). Re-performance. [Program note for Gershwin Plays]. Retrieved January 1, 2013 {but not May 18, 2022}, from http://www.zenph.com/gershwin/program-notes


NIGEL NETTHEIM has doctorates in Mathematical Statistics and in Musicology. Some of his research combines the two fields, and he has published widely in analytical musicology. A major publication was an annotated translation of Gustav Becking's How Musical Rhythm Reveals Human Attitudes (2011). He taught piano at the Royal Conservatory of Music, Toronto, and was a Fellow in the Music Research Centre at the Sydney Conservatorium of Music. He is an Adjunct Research Fellow at The MARCS Institute, University of Western Sydney {now Western Sydney University}, 2001 to date.


Nettheim home page                  Index of Pachmann items