Filter graph sample rate bug

[jnordberg]

Hi I'm trying to resample audio using filters but the resulting sample rate is 2x what I expect.

E.g. this gives me a slowed down version of the audio (96khz)

import av
import numpy as np
import torch
import torchaudio

container = av.open("audio_ref.wav")
stream = container.streams.audio[0]
print(stream)

target_sr = 48000

graph = av.filter.Graph()
graph.link_nodes(
    graph.add_abuffer(template=stream),
    graph.add("aresample", str(target_sr)),
    graph.add("abuffersink"),
).configure()

audio_frames = []
for frame in container.decode(audio=0):
    graph.push(frame)
    try:
        frame = graph.pull()
    except av.error.BlockingIOError:
        continue
    audio_frames.append(frame.to_ndarray())

container.close()

audio = torch.from_numpy(np.concatenate(audio_frames, axis=1))
torchaudio.save("audio_resampled.wav", audio, target_sr)

I found that if I divide the target sample rate by number of channels the output is correct

import av
import numpy as np
import torch
import torchaudio

container = av.open("audio_ref.wav")
stream = container.streams.audio[0]
print(stream)

target_sr = 48000

graph = av.filter.Graph()
graph.link_nodes(
    graph.add_abuffer(template=stream),
    graph.add("aresample", str(target_sr // stream.channels)), # why is this correct?
    graph.add("abuffersink"),
).configure()

audio_frames = []
for frame in container.decode(audio=0):
    graph.push(frame)
    try:
        frame = graph.pull()
    except av.error.BlockingIOError:
        continue
    audio_frames.append(frame.to_ndarray())

container.close()

audio = torch.from_numpy(np.concatenate(audio_frames, axis=1))
torchaudio.save("audio_resampled.wav", audio, target_sr)

Anyone know what's going on here?

[jnordberg]

I'm starting to think this is some bug.

Adding more filters changes the output, e.g. adding vibrato before or after aresample in the example above requires you to set sample rate without dividing by channels or the audio will be sped up.

Also this works as expected with same audio file:

ffmpeg -i audio_ref.wav -af "aresample=48000" -y audio_resampled.wav