❔ Transform bark (30 Mo) output wav to mp3/webm (pydub)

[adriens]

❔ About

i'm currently trying to compress a 30 Mo wav output, but pydub is always complaining about "error: Size should be 1, 2, 3 or 4".

👣 Steps to reproduce

Below is my code sample (I built the wav with bark ) :

!pip install pydub
!pip install ffmpeg-python

from pydub import AudioSegment
import ffmpeg


audio = AudioSegment.from_file('/kaggle/working/auptitcafe.wav', format='wav')
audio = audio.set_sample_width(2)
# Export the audio to MP3 format
audio.export('auptitcafe.mp3', format='mp3')

Then I get the following error message :

error                                     Traceback (most recent call last)
Cell In[22], line 9
      5 import ffmpeg
      8 audio = AudioSegment.from_file('/kaggle/working/auptitcafe.wav', format='wav')
----> 9 audio = audio.set_sample_width(2)
     10 # Export the audio to MP3 format
     11 audio.export('auptitcafe.mp3', format='mp3')

File /opt/conda/lib/python3.10/site-packages/pydub/audio_segment.py:1008, in AudioSegment.set_sample_width(self, sample_width)
   1003     return self
   1005 frame_width = self.channels * sample_width
   1007 return self._spawn(
-> 1008     audioop.lin2lin(self._data, self.sample_width, sample_width),
   1009     overrides={'sample_width': sample_width, 'frame_width': frame_width}
   1010 )

error: Size should be 1, 2, 3 or 4

[adriens]

Is there something special with bark's wav output ?

[C0untFloyd]

I'm unfamiliar with that module but Bark's Output Format is Mono, so perhaps try:
audio.set_channels(1)

I guess it wouldn't hurt to also specify the samplerate (which is 24000).

[adriens]

Hmmmn yes, thanks, i'm giving it a try and will let you know on this issue 🙏

[dnrico1]

Are you perhaps adding a silence as per the longform audio generation notebook? Ran into the same error yesterday when doing that, because np.zeroes produced 64-bit integers (which are like 64-bit wav samples). The error says that audioop only expects wav samples with 1, 2, 3 or 4 bytes per samples which are equivalent to 8, 16, 24 or 32-bit wavs. Inserting a 64-bit silence in the middle messes things up.

The solution was to specify dtype 16bit int for the np.zeroes

[adriens]

Are you perhaps adding a silence as per the longform audio generation notebook? R

Yes, indeed, this is how I'm adding silents. 😅

The solution was to specify dtype 16bit int for the np.zeroes

Would you share some code snippets ❔ 🙏

[C0untFloyd]

Not the poster above but you would just specify it in np.zeroes like this:
dtype=np.int16

[dnrico1]

Yes exactly dtype=np.int16 as an argument to np.zeroes

…

On Tue, 16. May 2023 at 20:45, C0untFloyd ***@***.***> wrote: Not the poster above but you would just specify it in np.zeroes like this: dtype=np.int16 — Reply to this email directly, view it on GitHub <#288 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A4HHAVFHMPCERF2RJO4ZWXTXGPDM3ANCNFSM6AAAAAAYAK4Q6U> . You are receiving this because you commented.Message ID: ***@***.***>

[adriens]

Thanks a lot to both of you @C0untFloyd and @dnrico1 , it really did the trick... making it possible to get the wav file into much much smaller files without hearable quality loss ❣️

# Properly code slience
# https://github.com/suno-ai/bark/issues/288
silence = np.zeros(int(0.25 * SAMPLE_RATE), dtype=np.int16 )

format	size
.wav	20.2 MB
.mp3	1.7 MB
.webm	1.8 MB

[adriens]

👇

• 🤏 Better silence management/encoding #305