Fans of classical music know that there are great performers and there are the rest. But those with any experience listening to music also know that a lot of what you hear depends on context: if you think you’re hearing a great performer, you’re likely to appreciate their performance even more.
In addition to my activities writing about Macs and iPods, I also review CDs for MusicWeb, an independent British web site that publishes reviews of classical CDs. So I was quite amused this morning when I read an article on the web site of the British classical music magazine Gramophone, entitled Masterpieces Or Fakes? The Joyce Hatto Scandal.
Joyce Hatto is a late pianist who had recorded a handful of good discs then suffered from cancer, from which she died in June 2006. Somewhere in the past year of her life, recordings started spewing out from a small label, run by her husband, showing this woman to have a surprising range of talents, and Gramophone, along with other publications, began championing these recordings. But some months ago, posters to the rec.music.classical.recordings newsgroup began questioning the possibility that this woman could have played all these works, with styles that sounded so different. One poster said:
“After hearing so much about Joyce Hatto, I started purchasing some of her recordings. While nothing I have heard is bad (in fact, I am glad I bought these CDs), I have noticed something eerie: that the pianist playing the Mozart sonatas cannot be the pianist playing Prokofiev
or the pianist playing Albeniz. I have the distinct feeling of being the victim of some sort of hoax. Does anyone else share these feelings?”
Well, where I come from, you might say, “them’s fightin’ words”, but they incited some people to start looking more closely at this phenomenon. The results seem to be clear (as shown in the Gramophone article linked above): not only was this a hoax, but a purely monetarily-driven one, which simply took copies of some works, fiddled with others, and released them to a world of people who fawn after the latest sensation.All this raises many questions, of course. First, you have to feel bad for the professional critics who, hearing something they liked, not only lauded it, but created the context to fulfill their wishes with each subsequent recording from this pianist. Second, it shows that there are, perhaps, some recordings by lesser-known musicians that had been “pirated” and branded with the Joyce Hatto name which merit further attention. Had these same critics panned the discs that were the actual sources of the Hatto recordings?
Finally, and perhaps more important, it shows the futility of any kind of criticism. Well, you can’t copy books or movies, but for classical music where critics review not so much the music as the interpretation and performance, how much criticism is truly objective? Perhaps it is time for critics to work blindly, getting nothing but blank discs (or digital files) and reviewing these, then, only after the reviews are filed, finding out who the performers are. This would, of course, not be to the liking of the major record labels, for whom marketing is often more important than actual performances. (Granted, this is only really valid for instrumental performances; it is relatively simple to recognize a familiar voice in an opera or other vocal recording.)
There has always been criticism of critics, but nowhere other than the classical music arena does the concept of “great performances” or “reference performances” hold sway. These are the benchmarks against which other performances are measured, and they can be self-fulfilling: the more familiar you are with your benchmark, the more you will like it and reinforce its validity.
I tend to be somewhat obsessive about music, and, for some composers, actively seek out different versions of works I like in order to have a variety of performances, because no one performance can be considered final or perfect. I have never succumbed to unfailing appreciation for a specific artist (though Dietrich Fischer-Dieskau is by far my preferred singer of German lieder), and tend to search broadly. It’s a shame for those who do think they found the new musical messiah, and many music publications–including Gramophone, who will have to make a serious mea culpa–will suffer from this type of hoax.
While it’s almost surprising this hasn’t happened before, there are actually a few reasons that make this case different from others. Joyce Hatto had not performed anywhere for a long time, so no one would have been able to compare her performance style with her recordings. Also, this was a very small label, and, while Hatto-mania may have blossomed, it certainly never went far enough to generate large sales. It seems that the greed behind this hoax was limited to a single person, the late pianist’s husband.
But with digital technology so prevalent, such that anyone can copy a CD and release it as their own, no one has time to check all the recordings that are released to make sure they are what they say. (Kudos to Andrew Rose of Pristine Audio for taking the time to analyze these recordings down to their waveforms; check this link for examples, both audio and visual, proving that the Hatto recordings are not indeed Hatto recordings.) While it is unlikely that there are many unscrupulous record labels who would consider perpetrating such a hoax, the cat’s out of the bag, and this may give ideas to others. Caveat emptor, right?
I’ve never listened to a Joyce Hatto recording. And this may very well be a
hoax — it probably is. However, I’d suggest caution rather than rushing to
judgement, because a good deal of the comment regarding this scandal
seems to be technically ill-informed.
The most persuasive case is made here:
<http://www.charm.rhul.ac.uk/content/contact/hatto_article.html>
But other articles (Gramophone, NYT, IHT, Classics Today, etc) are very short
or misleading on crucial technical elements. For instance, the scandal was
supposed to have started with a critic and former Hatto fan inserting one of
her CDs in his computer and having iTunes identify the performer as Laszlo
Simon instead of Joyce Hatto. But anyone who understands how iTunes
identifies audio CD tracks, and is familiar with the character of the Gracenote
CDDB and its limitations with respect to classical music, will immediately
realise that iTunes misidentifying the performer is proof of nothing.
The issue becomes even more complicated when we find out that, according
to Andrew Rose <http://www.pristineclassical.com/HattoHoax.html>,
of the twelve tracks on the Hatto CD, ten were actually by Laszlo Simon, and
two by Minoru Nojima. Such an audio CD should not have been identified by
iTunes either as a performance by Simon or one by Nojima. Moreover,
Andrew Rose argues that the Simon recordings attributed to Hatto were time-
shrunk by 0.02%, which would have altered the track duration enough, once
more, to prevent iTunes from recognising the CD as one by Laszlo Simon.
Gramophone states that Andrew Rose "scientifically checked the sound-
waves"; but a visual comparison in Adobe Audition of two different sound-
waves of the same music is anything but "scientific".
It is entirely possible, perhaps even likely, that all this is true: Joyce Hatto’s
recordings were fakes, Andrew Rose’s comparisons were valid, and so on. But
the contradictions and muddled (or lack of) methodological and technical
details so far should give one pause. It seems to me that, just as critics
rushed to proclaim Joyce Hatto an unrecognised genius, they are now rushing
to proclaim her a hoax, without really understanding the technology involved.
These critics need not only to update their skills for the digital age, but also
to reflect a little on Webster’s definition of "critical": "[…] to exercise careful
judgement and selection […]".
I’ve asked the same question about the iTunes thing, because I know how it
works with Gracenote – it compares frames and offsets, so anything that is
not _exactly_ the same will not turn up a disc which is a (partial) copy.
However, this is the disc it turned up:
http://www.gracenote.com/music/album.html/genclassical/
865115b0574fa1ba9388bfa349a9038e.html
Laszlo Simon / Franz Liszt -12 Etudes d’execution transcendante
Label: Concert Artists
My thought is that someone who was in on the scam got this info onto the
Gracenote CDDB in the hopes that it would eventually be discovered; if the
performer were listed as Hatto, or the label as Bis, then I would accept some
sort of (wildly improbable) coincidence, but this looks too much like a plant.
Second, regarding Andrew Rose, he’s a very talented engineer, having worked
with the BBC, and working now restoring old recordings. You can listen to
exceprts on his web site – it’s obvious that he discovered the "originals" for
the Hatto tracks he has examined.
But which disc is it? The Simon CD, with the wrong label? The Hatto CD with
the wrong artist? (Remember, according to Rose, two of ten tracks are
Nojima’s, not Simon’s.) And, when iTunes came up with "Simon" as the
performer on the Hatto CD, why did the critic mutter darkly, "Something is
rotten in the state of Denmark", instead of dismissing it as yet another
example of the countless errors in the classical music section of CDDB? And
why is it that none of the critics and bloggers thought to ask such basic
questions?
As to Andrew Rose, the issue is neither his expertise nor his probity (both of
which are beyond question). The issue is the description of his conclusions as
"scientific" (Gramophone, NYT, etc.) Visual examination of two waveforms is
no more scientific than a critic listening to two different recordings and
pronouncing them to be of the same performer.
The CHARM article is closer to the "scientific" label (and thus more
persuasive), but it, too, falls short. It contains no reference and no description
of the methodology of building what they call "timescapes"; where they use a
well-established parameter (the Pearson correlation coefficient), they do not
explain what data was used to calculate it and how it was acquired. Worse,
some of the results seem to be contradictory: the Indjic/Rubinstein 1939
comparison yields r < 0.8 for Opp. 17 and 68, but an amazingly high value
(0.954) for the "entire set".
Again, why is it that none of the critics/bloggers have raised these issues?
My guess is that someone who had info about the scam managed to get the
disc’s info on the Gracenote CDDB. It is not Gracenote that spotted the
performer as being Simon, but rather the disc as matching a record in its
database. So someone had to add that record (which lists Simon as the
performer of the Hatto disc, and not on Bis but on Concert Artists). I don’t
know if it’s possible to find out when that info was added. But it was clearly a
clue planted by someone.
Why haven’t others examined this? Because they don’t understand it. They
don’t realize that the Gracenote CDDB merely matches frames and offsets to
find a record that corresponds to a disc ID it uses. It doesn’t "analyze" the
music (as a New Scientist article suggests). Most journalists are merely
reporting what they’ve read elsewhere. However, the Gramophone article is
wrong, when it says that Distler discovered this. Actually, it was a reader of
Classics Today who sent the info to that site (where Distler writes reviews)
who then passed the info on to Gramophone who went public before Classics
Today. Sigh.
As for scientific, I think that waveforms are pretty unique; you’d have a match
for a short passage, perhaps, but for entire pieces? Doubtful. It will most
likely be the courts that decide which approach is sufficient to inculpate
Barrington-Coupe, but in the mean time it’s interesting to see how more and
more discs are indentified, and how little-known pianists will benefit from
this.
"My guess is that someone…" Perhaps it was a Freudian slip of Barrington-
Coupe’s. Or perhaps it was Barrington-Coupe, racked by guilt and remorse,
driven by an unconscious urge to confess. Or — and I find this most likely —
it was an Albanian dwarf named Cyrano, kept chained to a dank wall in the
basement of the Concert Artist studio. At any rate, we can guess as much as
we like, but guesses aren’t proof of anything, especially not of journalists and
music critics doing the jobs they’re paid for.
"Most journalists are merely reporting what they’ve read elsewhere." A telling
indictment of the profession, and the Hatto stories do much to prove it. No
wonder journalists rate almost as low as politicians in public confidence polls.
"As for scientific, I think that waveforms are pretty unique". The symptom of
the problem in a nutshell. It is not the OBJECT, but the METHOD which makes
a study scientific or not. Astronomy and astrology study the same objects —
one is a science, the other is not. Palaeontology and "creation science" study
the same objects — one is a science, the other is not. Waveforms being
"pretty unique" does not make a visual comparison of two waveforms
"scientific". As a matter of fact, if you look carefully at the Hatto and Simon
waveform pics posted by Rose, you will notice, even at low rez, that they are
not identical. Do the differences matter? Based on his experience, Rose says
they don’t, and I believe him; but that doesn’t make his opinion "scientific".
He may be wrong, and I may be wrong to believe him, just as wrong as the
music lovers who believed Gramophone when it proclaimed Joyce Hatto "the
greatest living pianist that no one has heard of" — the same Gramophone
who now proclaims Joyce Hatto a hoax.
This whole story shows that the journalists, the critics, the bloggers, the
pundits, etc, have learned absolutely nothing. Yesterday they rushed to
proclaim Hatto a genius — without even the minimal common sense of
asking how it was possible for a cancer-afflicted elderly woman who hadn’t
performed in public for two decades suddenly to start producing a massive
volume of recordings, all of top technical quality, covering the entire gamut
of the piano repertoire. Today, the same people, lemming-like, are rushing to
proclaim her a fraud — without even the minimal common sense to ask a few
basic questions. (How exactly did iTunes do what it is supposed to have
done? What exactly is a "scientific" study of waveforms? How valid is a
"scientific" study which posts a value indicating almost complete statistical
correlation between Eugen Indjic of 1988 and Rubinstein of 1939?)
The moral of the story is that journalists, music critics, bloggers, etc — all
those whom we assume to have a broader education, a finer sensibility, a
more discerning intellect, and a more inquiring mind than the rest of us hoi
polloi — turn out to be no more than a flock of sheep, baahing mindlessly
while they rush hither and yon as some bellwether leads them. The real
problem is that, after all, it’s us, not them, who end up being fleeced.
Because sensationalism sells, and rational thought puts the public
to sleep. Anyway, it is finally occurring here we noticed.
I am a co-author of the CHARM
report. We had been preparing the report as an article in response
to Gramophone’s March 2006 challenge for compelling evidence that Hatto
was not genuine; however, the scandal surfaced independently, so we just
deposited the current state of our report on our website, since problems
dealing with libel would be the responsibility of the news breaker. Thus,
a full analysis was not yet prepared in writing. That is our excuse for
it falling short. In any case, it was indended to only point out the
amazing conincidences between the two sets of performances, nothing else.
Why they are so similar is still a complete mystery to us…
Timescapes show correlations at different timescales, but they do not show
probabilities (they do show probability in a weak sense). So timescapes
can not be used to scientifically or legally prove that the two mazurka
sets were the same. The importance of timescapes is that they were
instrumental in the discovery of the striking similarity between the
two mazurka sets, and they can thus place the debate of similarity on
a level that even a three-year-old can comprehend (I am not sure about
adults, though). Examine the timescape pictures related to the report
which are found on the webpage:
http://mazurka.org.uk/ana/pcor
Look at the various pictures, and you will notice that most contain
structures that indicate either random or real similarities between
multiple performances of the same mazurka (there are two pairs of
timescapes where I can actually see that one performer owns the
recorings of another pianists on CD).
However, there are a few exceptions to the norm in the plots. This is
when the entire plot show a solid field of color. This indicates that
the best correlation at all timescales always matches to a single
other performance (out of typically 30 performances per work). This is
expected to occur when the plots represent different re-releases of the
same performance, but it is very extremely unlikely to occur between two
independent performances. For example, there are multiple Rubinstein
performances of the same mazurka on the timescape pages. He is most
similar to himself in other performances, but there are other closer
similarities that break the picture up into multiple regions.
The information is not contradictory. Correlation does not imply
causation. Or in other words, correlation does not imply probability.
For example, correlation between sets of pairs of numbers can be very
high just due to random processes.
What you wanted to see in the article are probabilities, which cannot
really be calculated from the correlations. However, we have calculated
probabilities related to the Hatto/Indjic mazurka sets which are
outlined below:
Probability that Hatto track times matched Indjic track times by
random coincidence. All track times match between the two within about +/- 5
seconds (due to time stretching of one of the sets and different amounts of silence padding). Assume that
the probability that any one mazurka performance can be matched at random
to within 5 seconds of another independent performance is about 75% likely. This is a very
rough assumption, but it should be fair and conservative. Therefore the
probability that all 49 mazurkas which I was interested in (there are
about 10 more on the two disks) can be calculated by this equation:
Thus, the probability that the the track times match so closely is
approximately one in a million. This cannot be caluculated from the
r-value for the correlation between Indjic and Hatto track times directly. This is not scientific proof that the two
recordings are the same, and it is not legal-level proof either, since
one in a million does still can leave a reasonable amount of doubt.
Probability that a single Hatto recording is the same as the matching
Indjic recording. I use Mazurka in A minor Op. 17, No. 4 as an example
in this case. I measured the downbeat locations in an independent
manner. The accuracy of my timing measurement is 10 milliseconds.
I then subtracted the difference between the two beat-time sets.
There was a linear increase in the differences due to time shifting in
one of the performances. This was removed. What resulted was a plot
of timing differences between about 400 beats in the piece. There was a
region of beat timing differences which was about 200 beats long, where
the average beat timings were all less than 11 milliseconds. If a human
were to try to tap to a constant beat, they would be expected to get less
than 1/2 of the taps within 25 milliseconds (note that mazurkas are not
performed at a steady tempo, so it actually should be a larger deviation).
Being able to hit the beat timings at less than a 25 millisecond deviation
for 200 consecutive beats is equivalent to flipping a coin and having it
come up heads 200 consecutive times. The probability of that occuring
is one in 1^200 or 1 in 10^59.
This proves beyond a reasonable doubt that the two recordings are the
same. 1 in 10^59 is approximately the same as one atom of hydrogen out of
all other atoms of hydrogen in a star. And this is only the probability
for a single performance on the CDs. Additionaly, if you were to examine
the waveform for acoustic similarities (they have different durations
and filtering, so are not exact copies), the probability that they are
the same performance skyrockets much, much higher.
> The probability of that occuring is
> one in 1^200 or 1 in 10^59.
Ack, I meant:
1 in 2^200, or 1 in 10^59
1^200 is still 1. Anyway it was a typo, and I don’t want
anyone pointing out the numerous spelling mistakes I made…
> Thus, a full analysis was not yet prepared in writing. That is our
> excuse for it falling short.
That a "full analysis" hasn’t been prepared yet is understandable. What is not
understandable is the complete silence regarding data and methodology.
What exactly was the raw data used? How was it acquired and processed?
What was the algorithm employed to generate "timescapes"? This shouldn’t
require a lot of time to deal with: if it has been published all you need to do is
add the refs. If it’s a novel method, somebody’s on-going thesis work, then
this needs to be said, and, once again, it shouldn’t take too much time or
space to do so. Moreover, on your web page you don’t have the stylistic and
space constraints of a journal article, so you don’t have to worry about
formatting and you can post all the raw data you wish.
> The importance of timescapes is that they were instrumental in the
> discovery of the striking similarity between the two mazurka sets, and
> they can thus place the debate of similarity on a level that even a
> three-year-old can comprehend […]
The importance of "timescapes" cannot be determined without understanding
how they were generated, just as you cannot understand what an X-ray
shows if you don’t know how it was generated. I said (and I still believe) that
your article made the "most persuasive case", but my belief was based on a
fundamentally unscientific assumption. Without the process, timescapes are
really just pretty pictures, no doubt highly gratifying to a three-year-old, but
falling somewhat short of that for an adult with minimal scientific literacy.
> The information is not contradictory.
It may not be, but I need a better explanation for Indjic/Rubinstein 1939 r
values of 0.616 (Op. 17/4), 0.664 (Op. 68/3), and 0.954 (entire set).
> Correlation does not imply causation.
You cannot have it both ways. Either r says something, in which case you
have to explain not only the values for Hatto/Indjic, but also those for Indjic/
Rubinstein; or it does not, in which case, why was it featured at all? In fact,
one could make a good case that r should not have been included in the
paper to begin with, because we already know the variables are correlated —
they are performances of the same musical piece! But it was your choice to do
so.
> What you wanted to see in the article are probabilities […]
Actually, no, that was not what I wanted. To recapitulate, what I felt was
missing was, (a) methods of data acquisition and processing, (b) methods of
building timescapes, (c) discussion of apparent inconsistencies in presented
data, (d) tables of all values used to generate timescapes and calculate
Pearson’s coefficient.
> Moreover, on your web page you don’t have the
> stylistic and space constraints of a journal article,
> so you don’t have to worry about formatting and you
> can post all the raw data you wish.
The timescape pictures were generated from beat timing data found on
the webpage:
http://mazurka.org.uk/info/excel/beat
(In MS Excel format, and a simpler text file containing same data).
> if it has been published all you need to do is add the refs.
> If it’s a novel method, somebody’s on-going thesis work, then
> this needs to be said
(1) It has not been published yet :-), and (2) it is a novel method.
(3) William Barrington-Coupe has admitted to fraud, so no need to dot
anymore t’s or cross anymore eyes in the aborted article.
If you are interested, there is a description of a predecesor method
call "keyscapes" which I described two years ago in this article:
Sapp, Craig. "Visual Hierarchical Key Analysis" in Computers in
Entertainment [http://www.acm.org/pubs/cie.html], volume 3, no. 4
(Oct 2005)
Also an earlier description is found in the Proceedings of the
International Computer Music Conference 2001. I adapted the method for
performance analysis this summer with only one minor modification in
the methodology.
The application for performance analysis will be presented next month
in the CHARM Symposium 4: Methods for analysing recordings which will
be held just outside of London on 12-14 April:
http://www.charm.rhul.ac.uk/content/events/symp4.html
An article that I am writing on the method will appear in an issue of
Musicae Scientiae towards the end of this year:
http://musicweb.hmt-hannover.de/escom/english/MusicScE/MSstart.htm
Also, I will probably submit an article related to timescapes to the
ISMIR 2007 conference, so you might be able to find it on their website
in the Fall: http://ismir2007.ismir.net
—————————————————————–
Timescape methodology in brief: Take a 1-dimensional string of numbers
(beat tempo or beat durations in this case) and perform pair-wise
correlation (Pearson) between a reference performance and the other
performances’s data. The correlations are done in increasingly larger
sub-sequences of data, until the entire data set pairs are correlated to
each other.
For example, if there are 6 numbers (A,B,C,D,E,F) in the performance
sequence, these sub-sequence pairings are used to calculate correlation
values:
AB, BC, CD, DE, EF, ABC, BCD, CDE, DEF, ABCD, BCDE, CDEF, ABCDE,
BCDEF, ABCDEF
These correlations are done between the reference performance and the
other performances being compared to the reference performance.
Next, a 2-dimensional plotting diagram is filled from small-scale
subsequences at the bottom to the complete sequence correlations at
the top. For the example sequence, this is approximately how the 2d
plotting area is arranged:
ABCDEF
ABCDE, BCDEF
ABCD, BCDE, CDEF
ABC, BCD, CDE, DEF
AB, BC, CD, DE, EF
(Imagine each line being centered, since the spacing is
altered due to web formatting)
For correlation, the sub-sequences A, B, C, D, E, and F are not
interesting, but for other sub-sequence calculations (such as average
tempo instead of correlation), I might add them to the bottom row.
Now, each cell in the plotting area is assigned a color. The color
is given by which other performance has the highest correlation value
to the reference performance. For example, if there were 100 performances
and 1 of them is the reference performance for the plot, then there are
99 other performances that were correlated with the reference at
sub-sequence BCDE. I would look at all 99 of those correlation
values and place a color in the BCDE slot in the timescape which
represents that performance.
In the Hatto/Indjic case, nearly all sub-sequences of their independent
beat duration/tempos correlated better than any of the other 30
or so performances by other pianists. The only ones which didn’t
match were the short sub-sequence correlations at the bottom of the
pictures which can randomly generate high correlation values by chance
since they are short sequences:
http://mazurka.org.uk/ana/pcor
Also, for those interested, I have generated "dynascapes" this week
which compare beat loudness (dynamics) between performances:
http://mazurka.org.uk/ana/pcor-gbdyn
Source data for these plots are found at:
http://mazurka.org.uk/info/excel/dyn/gbdyn
My focus group of three-year-olds can still pick out the Hatto/Indjic
pairing in these pictures :-).
Various slide shows which include some nicer figures can be found at
http://mazurka.org.uk/info/present
> timescapes are really … pretty pictures
Thanks, I agree 🙂
> > Correlation does not imply causation.
There can be many reasons why the r-value is high, and it is all the
more confusing since they are related to the same composition, as
you point out. The probability calculations give in my previous entry
are a more stronger basis for demonstrating plagiarism, since one of
them demonstrates that it is not humanly possible to imitate another
performance to such a high degree of accuracy. In other words, p-value
calculations are difficult for this type of performance analysis, and
I am interested in anyone who has suggestions…
> (a) methods of data acquisition and processing,
The data acquisition methodology is also novel and would take
too long to describe here in the margins :-), but basically
I am writing my own plugins to this free audio analysis editor:
http://www.sonicvisualiser.org
My plugins:
http://sv.mazurka.org.uk/download
If you own a copy of Chopin mazurkas for which I have extracted
data, you can load one of those performances into the editor,
and then download one of the annotation layer files from
the webpage (Sonic Visualiser can load a file directly
from a web address):
http://mazurka.org.uk/ana/markup
Such as the beat times in Joyce Hatto’s performance of Mazurka 17/4:
http://mazurka.org.uk/ana/markup/pid9073-15/contents.html
http://mazurka.org.uk/ana/markup/pid9073-15/pid9073-15-avgtap-sv.txt
I’ve never listened to a Joyce Hatto recording. And this may very well be a
hoax — it probably is. However, I’d suggest caution rather than rushing to
judgement, because a good deal of the comment regarding this scandal
seems to be technically ill-informed.
The most persuasive case is made here:
<http://www.charm.rhul.ac.uk/content/contact/hatto_article.html>
But other articles (Gramophone, NYT, IHT, Classics Today, etc) are very short
or misleading on crucial technical elements. For instance, the scandal was
supposed to have started with a critic and former Hatto fan inserting one of
her CDs in his computer and having iTunes identify the performer as Laszlo
Simon instead of Joyce Hatto. But anyone who understands how iTunes
identifies audio CD tracks, and is familiar with the character of the Gracenote
CDDB and its limitations with respect to classical music, will immediately
realise that iTunes misidentifying the performer is proof of nothing.
The issue becomes even more complicated when we find out that, according
to Andrew Rose <http://www.pristineclassical.com/HattoHoax.html>,
of the twelve tracks on the Hatto CD, ten were actually by Laszlo Simon, and
two by Minoru Nojima. Such an audio CD should not have been identified by
iTunes either as a performance by Simon or one by Nojima. Moreover,
Andrew Rose argues that the Simon recordings attributed to Hatto were time-
shrunk by 0.02%, which would have altered the track duration enough, once
more, to prevent iTunes from recognising the CD as one by Laszlo Simon.
Gramophone states that Andrew Rose "scientifically checked the sound-
waves"; but a visual comparison in Adobe Audition of two different sound-
waves of the same music is anything but "scientific".
It is entirely possible, perhaps even likely, that all this is true: Joyce Hatto’s
recordings were fakes, Andrew Rose’s comparisons were valid, and so on. But
the contradictions and muddled (or lack of) methodological and technical
details so far should give one pause. It seems to me that, just as critics
rushed to proclaim Joyce Hatto an unrecognised genius, they are now rushing
to proclaim her a hoax, without really understanding the technology involved.
These critics need not only to update their skills for the digital age, but also
to reflect a little on Webster’s definition of "critical": "[…] to exercise careful
judgement and selection […]".
I’ve asked the same question about the iTunes thing, because I know how it
works with Gracenote – it compares frames and offsets, so anything that is
not _exactly_ the same will not turn up a disc which is a (partial) copy.
However, this is the disc it turned up:
http://www.gracenote.com/music/album.html/genclassical/
865115b0574fa1ba9388bfa349a9038e.html
Laszlo Simon / Franz Liszt -12 Etudes d’execution transcendante
Label: Concert Artists
My thought is that someone who was in on the scam got this info onto the
Gracenote CDDB in the hopes that it would eventually be discovered; if the
performer were listed as Hatto, or the label as Bis, then I would accept some
sort of (wildly improbable) coincidence, but this looks too much like a plant.
Second, regarding Andrew Rose, he’s a very talented engineer, having worked
with the BBC, and working now restoring old recordings. You can listen to
exceprts on his web site – it’s obvious that he discovered the "originals" for
the Hatto tracks he has examined.
But which disc is it? The Simon CD, with the wrong label? The Hatto CD with
the wrong artist? (Remember, according to Rose, two of ten tracks are
Nojima’s, not Simon’s.) And, when iTunes came up with "Simon" as the
performer on the Hatto CD, why did the critic mutter darkly, "Something is
rotten in the state of Denmark", instead of dismissing it as yet another
example of the countless errors in the classical music section of CDDB? And
why is it that none of the critics and bloggers thought to ask such basic
questions?
As to Andrew Rose, the issue is neither his expertise nor his probity (both of
which are beyond question). The issue is the description of his conclusions as
"scientific" (Gramophone, NYT, etc.) Visual examination of two waveforms is
no more scientific than a critic listening to two different recordings and
pronouncing them to be of the same performer.
The CHARM article is closer to the "scientific" label (and thus more
persuasive), but it, too, falls short. It contains no reference and no description
of the methodology of building what they call "timescapes"; where they use a
well-established parameter (the Pearson correlation coefficient), they do not
explain what data was used to calculate it and how it was acquired. Worse,
some of the results seem to be contradictory: the Indjic/Rubinstein 1939
comparison yields r < 0.8 for Opp. 17 and 68, but an amazingly high value
(0.954) for the "entire set".
Again, why is it that none of the critics/bloggers have raised these issues?
My guess is that someone who had info about the scam managed to get the
disc’s info on the Gracenote CDDB. It is not Gracenote that spotted the
performer as being Simon, but rather the disc as matching a record in its
database. So someone had to add that record (which lists Simon as the
performer of the Hatto disc, and not on Bis but on Concert Artists). I don’t
know if it’s possible to find out when that info was added. But it was clearly a
clue planted by someone.
Why haven’t others examined this? Because they don’t understand it. They
don’t realize that the Gracenote CDDB merely matches frames and offsets to
find a record that corresponds to a disc ID it uses. It doesn’t "analyze" the
music (as a New Scientist article suggests). Most journalists are merely
reporting what they’ve read elsewhere. However, the Gramophone article is
wrong, when it says that Distler discovered this. Actually, it was a reader of
Classics Today who sent the info to that site (where Distler writes reviews)
who then passed the info on to Gramophone who went public before Classics
Today. Sigh.
As for scientific, I think that waveforms are pretty unique; you’d have a match
for a short passage, perhaps, but for entire pieces? Doubtful. It will most
likely be the courts that decide which approach is sufficient to inculpate
Barrington-Coupe, but in the mean time it’s interesting to see how more and
more discs are indentified, and how little-known pianists will benefit from
this.
"My guess is that someone…" Perhaps it was a Freudian slip of Barrington-
Coupe’s. Or perhaps it was Barrington-Coupe, racked by guilt and remorse,
driven by an unconscious urge to confess. Or — and I find this most likely —
it was an Albanian dwarf named Cyrano, kept chained to a dank wall in the
basement of the Concert Artist studio. At any rate, we can guess as much as
we like, but guesses aren’t proof of anything, especially not of journalists and
music critics doing the jobs they’re paid for.
"Most journalists are merely reporting what they’ve read elsewhere." A telling
indictment of the profession, and the Hatto stories do much to prove it. No
wonder journalists rate almost as low as politicians in public confidence polls.
"As for scientific, I think that waveforms are pretty unique". The symptom of
the problem in a nutshell. It is not the OBJECT, but the METHOD which makes
a study scientific or not. Astronomy and astrology study the same objects —
one is a science, the other is not. Palaeontology and "creation science" study
the same objects — one is a science, the other is not. Waveforms being
"pretty unique" does not make a visual comparison of two waveforms
"scientific". As a matter of fact, if you look carefully at the Hatto and Simon
waveform pics posted by Rose, you will notice, even at low rez, that they are
not identical. Do the differences matter? Based on his experience, Rose says
they don’t, and I believe him; but that doesn’t make his opinion "scientific".
He may be wrong, and I may be wrong to believe him, just as wrong as the
music lovers who believed Gramophone when it proclaimed Joyce Hatto "the
greatest living pianist that no one has heard of" — the same Gramophone
who now proclaims Joyce Hatto a hoax.
This whole story shows that the journalists, the critics, the bloggers, the
pundits, etc, have learned absolutely nothing. Yesterday they rushed to
proclaim Hatto a genius — without even the minimal common sense of
asking how it was possible for a cancer-afflicted elderly woman who hadn’t
performed in public for two decades suddenly to start producing a massive
volume of recordings, all of top technical quality, covering the entire gamut
of the piano repertoire. Today, the same people, lemming-like, are rushing to
proclaim her a fraud — without even the minimal common sense to ask a few
basic questions. (How exactly did iTunes do what it is supposed to have
done? What exactly is a "scientific" study of waveforms? How valid is a
"scientific" study which posts a value indicating almost complete statistical
correlation between Eugen Indjic of 1988 and Rubinstein of 1939?)
The moral of the story is that journalists, music critics, bloggers, etc — all
those whom we assume to have a broader education, a finer sensibility, a
more discerning intellect, and a more inquiring mind than the rest of us hoi
polloi — turn out to be no more than a flock of sheep, baahing mindlessly
while they rush hither and yon as some bellwether leads them. The real
problem is that, after all, it’s us, not them, who end up being fleeced.
Because sensationalism sells, and rational thought puts the public
to sleep. Anyway, it is finally occurring here we noticed.
I am a co-author of the CHARM
report. We had been preparing the report as an article in response
to Gramophone’s March 2006 challenge for compelling evidence that Hatto
was not genuine; however, the scandal surfaced independently, so we just
deposited the current state of our report on our website, since problems
dealing with libel would be the responsibility of the news breaker. Thus,
a full analysis was not yet prepared in writing. That is our excuse for
it falling short. In any case, it was indended to only point out the
amazing conincidences between the two sets of performances, nothing else.
Why they are so similar is still a complete mystery to us…
Timescapes show correlations at different timescales, but they do not show
probabilities (they do show probability in a weak sense). So timescapes
can not be used to scientifically or legally prove that the two mazurka
sets were the same. The importance of timescapes is that they were
instrumental in the discovery of the striking similarity between the
two mazurka sets, and they can thus place the debate of similarity on
a level that even a three-year-old can comprehend (I am not sure about
adults, though). Examine the timescape pictures related to the report
which are found on the webpage:
http://mazurka.org.uk/ana/pcor
Look at the various pictures, and you will notice that most contain
structures that indicate either random or real similarities between
multiple performances of the same mazurka (there are two pairs of
timescapes where I can actually see that one performer owns the
recorings of another pianists on CD).
However, there are a few exceptions to the norm in the plots. This is
when the entire plot show a solid field of color. This indicates that
the best correlation at all timescales always matches to a single
other performance (out of typically 30 performances per work). This is
expected to occur when the plots represent different re-releases of the
same performance, but it is very extremely unlikely to occur between two
independent performances. For example, there are multiple Rubinstein
performances of the same mazurka on the timescape pages. He is most
similar to himself in other performances, but there are other closer
similarities that break the picture up into multiple regions.
The information is not contradictory. Correlation does not imply
causation. Or in other words, correlation does not imply probability.
For example, correlation between sets of pairs of numbers can be very
high just due to random processes.
What you wanted to see in the article are probabilities, which cannot
really be calculated from the correlations. However, we have calculated
probabilities related to the Hatto/Indjic mazurka sets which are
outlined below:
Probability that Hatto track times matched Indjic track times by
random coincidence. All track times match between the two within about +/- 5
seconds (due to time stretching of one of the sets and different amounts of silence padding). Assume that
the probability that any one mazurka performance can be matched at random
to within 5 seconds of another independent performance is about 75% likely. This is a very
rough assumption, but it should be fair and conservative. Therefore the
probability that all 49 mazurkas which I was interested in (there are
about 10 more on the two disks) can be calculated by this equation:
Thus, the probability that the the track times match so closely is
approximately one in a million. This cannot be caluculated from the
r-value for the correlation between Indjic and Hatto track times directly. This is not scientific proof that the two
recordings are the same, and it is not legal-level proof either, since
one in a million does still can leave a reasonable amount of doubt.
Probability that a single Hatto recording is the same as the matching
Indjic recording. I use Mazurka in A minor Op. 17, No. 4 as an example
in this case. I measured the downbeat locations in an independent
manner. The accuracy of my timing measurement is 10 milliseconds.
I then subtracted the difference between the two beat-time sets.
There was a linear increase in the differences due to time shifting in
one of the performances. This was removed. What resulted was a plot
of timing differences between about 400 beats in the piece. There was a
region of beat timing differences which was about 200 beats long, where
the average beat timings were all less than 11 milliseconds. If a human
were to try to tap to a constant beat, they would be expected to get less
than 1/2 of the taps within 25 milliseconds (note that mazurkas are not
performed at a steady tempo, so it actually should be a larger deviation).
Being able to hit the beat timings at less than a 25 millisecond deviation
for 200 consecutive beats is equivalent to flipping a coin and having it
come up heads 200 consecutive times. The probability of that occuring
is one in 1^200 or 1 in 10^59.
This proves beyond a reasonable doubt that the two recordings are the
same. 1 in 10^59 is approximately the same as one atom of hydrogen out of
all other atoms of hydrogen in a star. And this is only the probability
for a single performance on the CDs. Additionaly, if you were to examine
the waveform for acoustic similarities (they have different durations
and filtering, so are not exact copies), the probability that they are
the same performance skyrockets much, much higher.
> The probability of that occuring is
> one in 1^200 or 1 in 10^59.
Ack, I meant:
1 in 2^200, or 1 in 10^59
1^200 is still 1. Anyway it was a typo, and I don’t want
anyone pointing out the numerous spelling mistakes I made…
> Thus, a full analysis was not yet prepared in writing. That is our
> excuse for it falling short.
That a "full analysis" hasn’t been prepared yet is understandable. What is not
understandable is the complete silence regarding data and methodology.
What exactly was the raw data used? How was it acquired and processed?
What was the algorithm employed to generate "timescapes"? This shouldn’t
require a lot of time to deal with: if it has been published all you need to do is
add the refs. If it’s a novel method, somebody’s on-going thesis work, then
this needs to be said, and, once again, it shouldn’t take too much time or
space to do so. Moreover, on your web page you don’t have the stylistic and
space constraints of a journal article, so you don’t have to worry about
formatting and you can post all the raw data you wish.
> The importance of timescapes is that they were instrumental in the
> discovery of the striking similarity between the two mazurka sets, and
> they can thus place the debate of similarity on a level that even a
> three-year-old can comprehend […]
The importance of "timescapes" cannot be determined without understanding
how they were generated, just as you cannot understand what an X-ray
shows if you don’t know how it was generated. I said (and I still believe) that
your article made the "most persuasive case", but my belief was based on a
fundamentally unscientific assumption. Without the process, timescapes are
really just pretty pictures, no doubt highly gratifying to a three-year-old, but
falling somewhat short of that for an adult with minimal scientific literacy.
> The information is not contradictory.
It may not be, but I need a better explanation for Indjic/Rubinstein 1939 r
values of 0.616 (Op. 17/4), 0.664 (Op. 68/3), and 0.954 (entire set).
> Correlation does not imply causation.
You cannot have it both ways. Either r says something, in which case you
have to explain not only the values for Hatto/Indjic, but also those for Indjic/
Rubinstein; or it does not, in which case, why was it featured at all? In fact,
one could make a good case that r should not have been included in the
paper to begin with, because we already know the variables are correlated —
they are performances of the same musical piece! But it was your choice to do
so.
> What you wanted to see in the article are probabilities […]
Actually, no, that was not what I wanted. To recapitulate, what I felt was
missing was, (a) methods of data acquisition and processing, (b) methods of
building timescapes, (c) discussion of apparent inconsistencies in presented
data, (d) tables of all values used to generate timescapes and calculate
Pearson’s coefficient.
> Moreover, on your web page you don’t have the
> stylistic and space constraints of a journal article,
> so you don’t have to worry about formatting and you
> can post all the raw data you wish.
The timescape pictures were generated from beat timing data found on
the webpage:
http://mazurka.org.uk/info/excel/beat
(In MS Excel format, and a simpler text file containing same data).
> if it has been published all you need to do is add the refs.
> If it’s a novel method, somebody’s on-going thesis work, then
> this needs to be said
(1) It has not been published yet :-), and (2) it is a novel method.
(3) William Barrington-Coupe has admitted to fraud, so no need to dot
anymore t’s or cross anymore eyes in the aborted article.
If you are interested, there is a description of a predecesor method
call "keyscapes" which I described two years ago in this article:
Sapp, Craig. "Visual Hierarchical Key Analysis" in Computers in
Entertainment [http://www.acm.org/pubs/cie.html], volume 3, no. 4
(Oct 2005)
Also an earlier description is found in the Proceedings of the
International Computer Music Conference 2001. I adapted the method for
performance analysis this summer with only one minor modification in
the methodology.
The application for performance analysis will be presented next month
in the CHARM Symposium 4: Methods for analysing recordings which will
be held just outside of London on 12-14 April:
http://www.charm.rhul.ac.uk/content/events/symp4.html
An article that I am writing on the method will appear in an issue of
Musicae Scientiae towards the end of this year:
http://musicweb.hmt-hannover.de/escom/english/MusicScE/MSstart.htm
Also, I will probably submit an article related to timescapes to the
ISMIR 2007 conference, so you might be able to find it on their website
in the Fall: http://ismir2007.ismir.net
—————————————————————–
Timescape methodology in brief: Take a 1-dimensional string of numbers
(beat tempo or beat durations in this case) and perform pair-wise
correlation (Pearson) between a reference performance and the other
performances’s data. The correlations are done in increasingly larger
sub-sequences of data, until the entire data set pairs are correlated to
each other.
For example, if there are 6 numbers (A,B,C,D,E,F) in the performance
sequence, these sub-sequence pairings are used to calculate correlation
values:
AB, BC, CD, DE, EF, ABC, BCD, CDE, DEF, ABCD, BCDE, CDEF, ABCDE,
BCDEF, ABCDEF
These correlations are done between the reference performance and the
other performances being compared to the reference performance.
Next, a 2-dimensional plotting diagram is filled from small-scale
subsequences at the bottom to the complete sequence correlations at
the top. For the example sequence, this is approximately how the 2d
plotting area is arranged:
ABCDEF
ABCDE, BCDEF
ABCD, BCDE, CDEF
ABC, BCD, CDE, DEF
AB, BC, CD, DE, EF
(Imagine each line being centered, since the spacing is
altered due to web formatting)
For correlation, the sub-sequences A, B, C, D, E, and F are not
interesting, but for other sub-sequence calculations (such as average
tempo instead of correlation), I might add them to the bottom row.
Now, each cell in the plotting area is assigned a color. The color
is given by which other performance has the highest correlation value
to the reference performance. For example, if there were 100 performances
and 1 of them is the reference performance for the plot, then there are
99 other performances that were correlated with the reference at
sub-sequence BCDE. I would look at all 99 of those correlation
values and place a color in the BCDE slot in the timescape which
represents that performance.
In the Hatto/Indjic case, nearly all sub-sequences of their independent
beat duration/tempos correlated better than any of the other 30
or so performances by other pianists. The only ones which didn’t
match were the short sub-sequence correlations at the bottom of the
pictures which can randomly generate high correlation values by chance
since they are short sequences:
http://mazurka.org.uk/ana/pcor
Also, for those interested, I have generated "dynascapes" this week
which compare beat loudness (dynamics) between performances:
http://mazurka.org.uk/ana/pcor-gbdyn
Source data for these plots are found at:
http://mazurka.org.uk/info/excel/dyn/gbdyn
My focus group of three-year-olds can still pick out the Hatto/Indjic
pairing in these pictures :-).
Various slide shows which include some nicer figures can be found at
http://mazurka.org.uk/info/present
> timescapes are really … pretty pictures
Thanks, I agree 🙂
> > Correlation does not imply causation.
There can be many reasons why the r-value is high, and it is all the
more confusing since they are related to the same composition, as
you point out. The probability calculations give in my previous entry
are a more stronger basis for demonstrating plagiarism, since one of
them demonstrates that it is not humanly possible to imitate another
performance to such a high degree of accuracy. In other words, p-value
calculations are difficult for this type of performance analysis, and
I am interested in anyone who has suggestions…
> (a) methods of data acquisition and processing,
The data acquisition methodology is also novel and would take
too long to describe here in the margins :-), but basically
I am writing my own plugins to this free audio analysis editor:
http://www.sonicvisualiser.org
My plugins:
http://sv.mazurka.org.uk/download
If you own a copy of Chopin mazurkas for which I have extracted
data, you can load one of those performances into the editor,
and then download one of the annotation layer files from
the webpage (Sonic Visualiser can load a file directly
from a web address):
http://mazurka.org.uk/ana/markup
Such as the beat times in Joyce Hatto’s performance of Mazurka 17/4:
http://mazurka.org.uk/ana/markup/pid9073-15/contents.html
http://mazurka.org.uk/ana/markup/pid9073-15/pid9073-15-avgtap-sv.txt