Re: [OT] Vivi, the Virtual Violinist, plays LilyPond music

[Janek Warchoł]

2011/3/5 Graham Percival <address@hidden>:
> On Sat, Mar 05, 2011 at 12:54:31AM +0100, Janek Warchoł wrote:
>> 2011/3/4 Graham Percival <address@hidden>
>> > Well, playing on a $1000 (or even $1,000,000) violin is just a
>> > matter of getting a recording of somebody tapping on such an
>> > instrument.  Such recordings (I only need 12 milliseconds of a tap
>> > noise!) _are_ available online, but I haven't yet found a
>> > recording which stated that it was available under the GPL.  And
>> > since sound recordings are covered by copyright, I can't just take
>> > an existing one.  :(
>>
>> Does this mean that Artifastring is already able to simulate violing
>> sounds so perfectly that everything is a matter of "teaching" it how
>> to do so?
>
> Artifastring is a very imperfect simulation of a violin.  You can
> think about this as having two stages: 1. the actions of the four
> strings, and 2. the actions of the body (the "big empty part" of
> the instrument).  The way that it simulates the effect of a violin
> body (or cello body) is by using a mathematical operation called
> "convolution".  This convolution is done by multiplying the sum of
> the output samples of the strings with literally an audio
> recording of tapping a violin.  In engineering terms, this is an
> approximation of the "impulse response".
>
> Because of this, switching to a different violin sound is purely a
> matter of switching the "tap" recording.
>
>> I don't have much experience with violin, but judging by the audio
>> samples i thought that it could use some improvement (independently of
>> Vivi's playing skill improvement).
>
> Yes and no.  The actions of the strings are imperfect, but the
> violin used for the impulse response really is a bad instrument.
> The "£100 pounds" figure actually included the bow, case, and
> shipping.
>
> Also, generating an impulse response by tapping the instrument is
> not a perfect impulse response.  It's a decent approximation, but
> serious acoustics researchers would use a frequency sweep or
> something like that instead.

Ah, now i understand. Thanks for explanation.

> But I'm not a serious acoustics researcher -- my goal is to
> advance the art of automatic music performance.  There's enough
> work that I can do on performing chords, vibrato, and the like,
> such that I'm not hugely concerned with the instrument quality at
> the moment.  And the methods for training Vivi are completely
> general; if/when the violin sounds change (or even changes to a
> cello or viola), all I need to do is spend 2-3 hours teaching her
> how to play that instrument, by classifying audio!  In other
> words, there's no programming involved in this training; I'm just
> acting like a parent of a music student.

Yes, thats probably the best part :)

> Of course, I'm hoping that when I present this at a conference,
> somebody from the audience will say "wow, that's nice work, but
> it's a pity that your physical model only approximates XYZ.  I
> have code that does this; could we work together?  I don't mind
> putting that code under the GPLv3."

:) That's roughly what i meant by $1000 violin :)

> So far, the only other open-source bowed-string simulation that I
> know of is the 1986 Smith algorithm, implemented in the Synthesis
> Toolkit in C++.  Most of the algorithms in Artifastring have been
> improved in the 2000s, but it's still much better than the version
> in STK.
> (the big open-source audio programs like Csound, Chuck, and
> supercollider, all use STK for physical modeling, which means the
> 25-year-old string algorithm)

Whoah! That's suprising.

cheers,
Janek