Problem is in conversion from float to short and int in libsndfile. Libsndfile's code don't care about normalizing floats when converts them to integer and we get samples "short = float * 32767" when float above 1.0 or below -1.0.

from libsndfile.

No effect with SFC_SET_SCALE_FLOAT_INT_READ on OGG files

from libsndfile.

Sorry, really busy this week. Its unlikely I'll have a chance to look at this before the weekend.

from libsndfile.

And sndfile-info:

sndfile-info st11.ogg

Version : libsndfile-1.0.26pre2

File : /media/166/st11.ogg
Length : 1818885
Ogg stream data : Vorbis
Stream serialno : 28657
Vorbis library version : Xiph.Org libVorbis 1.3.3
Bitstream is 2 channel, 44100 Hz
Encoded by : Xiph.Org libVorbis I 20020717
Metadata :
Title :
Artist :
End

Sample Rate : 44100
Frames : 4188708
Channels : 2
Format : 0x00200060
Sections : 1
Seekable : TRUE
Duration : 00:01:34.982
Signal Max : 1.18905 (-88.80 dB)

from libsndfile.

Sorry, really busy this week. Its unlikely I'll have a chance to look at this before the weekend.

No problem.

from libsndfile.

Patch for this issue http://pastebin.com/wME3zFRm or my pull request

from libsndfile.

Hi

Using sf_command(file, SFC_SET_SCALE_FLOAT_INT_READ, NULL, SF_TRUE) seems to take forever (15 seconds now, almost unnoticeable before).

Could it be because this function has to iterate on all the samples of the audio file? Is there an easy fix or workaround?

from libsndfile.

@LaurentGomila Yes, that is the case and no there is no easy workaround.

The fundamental problem is that reading floating point samples into short or int arrays is likely to go wrong unless you know beforehand the the maximum sample value.

The best way of avoiding the problem is to only read the data as float or double.

from libsndfile.

I see. Thank you for the fast answer.

from libsndfile.

In case SFC_SET_SCALE_FLOAT_INT_READ is not used, can't you at least clamp the float values to 1 instead of getting integers greater than 32767 that wrap to -32768 and produce cracks when the sound is played? Clamping those values should make the "errors" almost unnoticeable.

from libsndfile.

Your proposed solution works very poorly for sound files where the float values are in the range [-32768.0, 32767.0].

Files like this do exist in the wild.

from libsndfile.

Oh, how about make dynamic scaller flag? It's starts' from 1.0 and if find
greater value while reading - it will set new maximum. It's will not solve
problem with identic and quality sound but will solve choppy sound.

2012/10/27 Erik de Castro Lopo [email protected]

Your proposed solution works very poorly for sound files where the float
values are in the range [-32768.0, 32767.0].

Files like this do exist in the wild.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/16#issuecomment-9827957.

from libsndfile.

Your proposed solution works very poorly for sound files where the float values are in the range [-32768.0, 32767.0].

I see. But what happened for this kind of values with the previous code, where all values above 1 were overflowing the short integer capacity? It wasn't more correct, was it?

from libsndfile.