SndBuf - audio file with samplerate different from 44100
I forgot to say, that the program I posted in the previous mail returns a lot of errors, basically 97% of the file length. I suppose the samples which are the same are all 0. btw the bit depth is the same, they're both 16 bit. cheers, Mario
Hey Mario, Thanks so much for your work on the manual, its looking great! SndBuf/SndBuf2 are designed to resample the audio file to the native rate when doing audio playback, although off the top of my head I don't know if valueAt()/samples() are also resampling (seems like they shouldn't, to allow true sample-level access). Generally speaking, comparing two floats for exact equality is too rigorous for digital audio. Its preferred to test that they are "close enough" within a desired order of magnitude, e.g. Math.fabs(f1-f2) < Math.fabs(f1)*0.0001 (see e.g. [1]). Secondly, there are at least two resamplings involved in this test (when you created perc2.wav, it was resampled from perc1.wav at 44100 to 48000, and then ChucK might be resampling it back to 44100). Under certain conditions resampling can be theoretically "perfect," but otherwise its just making a guess what the sample would be at the new rate. Even under perfect conditions, the inexact nature of floating point arithmetic means that resampling from 44100 -> 48000 -> 44100 will most likely result in a different series of actual values. Spencer [1] http://floating-point-gui.de/errors/comparison/ On Sat, Mar 24, 2018 at 8:59 AM, mario buoninfante < mario.buoninfante@gmail.com> wrote:
I forgot to say, that the program I posted in the previous mail returns a lot of errors, basically 97% of the file length. I suppose the samples which are the same are all 0. btw the bit depth is the same, they're both 16 bit.
cheers,
Mario
_______________________________________________ chuck-users mailing list chuck-users@lists.cs.princeton.edu https://lists.cs.princeton.edu/mailman/listinfo/chuck-users
-- Spencer Salazar, PhD Special Faculty Music Technology: Interaction, Intelligence, and Design California Institute of the Arts ssalazar@calarts.edu | +1 831.277.4654 https://spencersalazar.com
Hi Spencer,
thanks for your help, you're perfectly right about the floating point
comparison, I didn't think about it. and I think you're also right when you
say that valueAt() and samples() are ignoring the sample rate conversion
made by SndBuf. What I didn't say in the previous mail is that the way I
discovered this discrepancy is when I transferred all the samples from
SndBuf to Wavetable (chugin). basically I load an audio file in SndBuf then
read trough it using valueAt() and copy all the samples into an array
(array length set using .sample() ). then this array is used with
Wavetable. I noticed that something was wrong when I played Wavetable with
a Phasor and the pitch was wrong. only at that point I ran the test where I
compared the 2 two SnbBuf .valueAt().
Btw later I'll have another look at .valueAt()/.samples() and try to
figure out whether they consider the sample rate conversion or not.
cheers,
Mario
On Sat, Mar 24, 2018 at 5:11 PM, Spencer Salazar
Hey Mario,
Thanks so much for your work on the manual, its looking great!
SndBuf/SndBuf2 are designed to resample the audio file to the native rate when doing audio playback, although off the top of my head I don't know if valueAt()/samples() are also resampling (seems like they shouldn't, to allow true sample-level access).
Generally speaking, comparing two floats for exact equality is too rigorous for digital audio. Its preferred to test that they are "close enough" within a desired order of magnitude, e.g. Math.fabs(f1-f2) < Math.fabs(f1)*0.0001 (see e.g. [1]).
Secondly, there are at least two resamplings involved in this test (when you created perc2.wav, it was resampled from perc1.wav at 44100 to 48000, and then ChucK might be resampling it back to 44100). Under certain conditions resampling can be theoretically "perfect," but otherwise its just making a guess what the sample would be at the new rate. Even under perfect conditions, the inexact nature of floating point arithmetic means that resampling from 44100 -> 48000 -> 44100 will most likely result in a different series of actual values.
Spencer
[1] http://floating-point-gui.de/errors/comparison/
On Sat, Mar 24, 2018 at 8:59 AM, mario buoninfante < mario.buoninfante@gmail.com> wrote:
I forgot to say, that the program I posted in the previous mail returns a lot of errors, basically 97% of the file length. I suppose the samples which are the same are all 0. btw the bit depth is the same, they're both 16 bit.
cheers,
Mario
_______________________________________________ chuck-users mailing list chuck-users@lists.cs.princeton.edu https://lists.cs.princeton.edu/mailman/listinfo/chuck-users
-- Spencer Salazar, PhD Special Faculty Music Technology: Interaction, Intelligence, and Design California Institute of the Arts
ssalazar@calarts.edu | +1 831.277.4654 <(831)%20277-4654> https://spencersalazar.com
_______________________________________________ chuck-users mailing list chuck-users@lists.cs.princeton.edu https://lists.cs.princeton.edu/mailman/listinfo/chuck-users
Hi, a quick update, basically there are a couple of things I missed before. I was using a stereo file in a mono SndBuf and the Phasor I was using to drive the Wavetable (I was copying the audio file into a Wavetable) needed twice the normal freq. it seems like loading a stereo file in a mono SndBuf means filling an array with samples form both left and right channel, something like: sndbuf = [1_left, 1_right, 2_left, 2_right, 3_left, 3_right, ...] - numbers indicate sample number of course I'm not sure about that, but this will justify the fact that Phasor needs a freq twice faster then the expected one when I load a stereo file, and also the fact that the sample length is half the original. apart from that, I think .valueAt() and .sample() both access to the raw data, as expected. basically I think there's no issue with SndBuf at all, it was me not considering all this obvious thing when I used Phasor to read the file, and the issue I was experiencing with this can be fixed simply multiplying phasor frequency by (audio file sample rate / ChucK sample rate). ie to read the entire file at the original speed: phasor.freq( (1000*(audioFileSR/ChucK_SR)) / (sample2.samples()::samp/ms) ) - sorry for the terrible syntax, it's just to show the logic behind. all this makes me think that it would be great having a method that returns the sample rate of the audio file loaded. :) cheers, Mario cheers, Mario On Sat, Mar 24, 2018 at 5:29 PM, Mario Buoninfante < mario.buoninfante@gmail.com> wrote:
Hi Spencer,
thanks for your help, you're perfectly right about the floating point comparison, I didn't think about it. and I think you're also right when you say that valueAt() and samples() are ignoring the sample rate conversion made by SndBuf. What I didn't say in the previous mail is that the way I discovered this discrepancy is when I transferred all the samples from SndBuf to Wavetable (chugin). basically I load an audio file in SndBuf then read trough it using valueAt() and copy all the samples into an array (array length set using .sample() ). then this array is used with Wavetable. I noticed that something was wrong when I played Wavetable with a Phasor and the pitch was wrong. only at that point I ran the test where I compared the 2 two SnbBuf .valueAt(). Btw later I'll have another look at .valueAt()/.samples() and try to figure out whether they consider the sample rate conversion or not.
cheers, Mario
On Sat, Mar 24, 2018 at 5:11 PM, Spencer Salazar < spencer.salazar@gmail.com> wrote:
Hey Mario,
Thanks so much for your work on the manual, its looking great!
SndBuf/SndBuf2 are designed to resample the audio file to the native rate when doing audio playback, although off the top of my head I don't know if valueAt()/samples() are also resampling (seems like they shouldn't, to allow true sample-level access).
Generally speaking, comparing two floats for exact equality is too rigorous for digital audio. Its preferred to test that they are "close enough" within a desired order of magnitude, e.g. Math.fabs(f1-f2) < Math.fabs(f1)*0.0001 (see e.g. [1]).
Secondly, there are at least two resamplings involved in this test (when you created perc2.wav, it was resampled from perc1.wav at 44100 to 48000, and then ChucK might be resampling it back to 44100). Under certain conditions resampling can be theoretically "perfect," but otherwise its just making a guess what the sample would be at the new rate. Even under perfect conditions, the inexact nature of floating point arithmetic means that resampling from 44100 -> 48000 -> 44100 will most likely result in a different series of actual values.
Spencer
[1] http://floating-point-gui.de/errors/comparison/
On Sat, Mar 24, 2018 at 8:59 AM, mario buoninfante < mario.buoninfante@gmail.com> wrote:
I forgot to say, that the program I posted in the previous mail returns a lot of errors, basically 97% of the file length. I suppose the samples which are the same are all 0. btw the bit depth is the same, they're both 16 bit.
cheers,
Mario
_______________________________________________ chuck-users mailing list chuck-users@lists.cs.princeton.edu https://lists.cs.princeton.edu/mailman/listinfo/chuck-users
-- Spencer Salazar, PhD Special Faculty Music Technology: Interaction, Intelligence, and Design California Institute of the Arts
ssalazar@calarts.edu | +1 831.277.4654 <(831)%20277-4654> https://spencersalazar.com
_______________________________________________ chuck-users mailing list chuck-users@lists.cs.princeton.edu https://lists.cs.princeton.edu/mailman/listinfo/chuck-users
participants (3)
-
mario buoninfante -
Mario Buoninfante
-
Spencer Salazar