PyTorch Split a Wave to Some Clips with Same Length – PyTorch Tutorial

By | January 11, 2024
0 Comment

In this tutorial, we will use an example to show you how to split a big wave file to some clips (small wave files) with same length. For example, there is a wave file with 500 seconds, we plan to split it to some small files, the duration of each small wave file is 5 senconds. How to do?

In pytorch, we can use the example code below:

  1. import torch
  2. import torchaudio
  4. def load_audio(wave_file):
  5. wave_data, sr = torchaudio.load(wave_file)
  6. return wave_data, sr
  8. def save_audio(wave_data, wave_file):
  9. torch.save(wave_data.clone(), wave_file, _use_new_zipfile_serialization=True)
  10. def get_audio_clips(wave_file, min_sec = 5):
  11. wave_data, sr = load_audio(wave_file)
  13. min_clip_len = min_sec * sr
  15. wave_len = wave_data.shape[1]
  16. if wave_len < min_clip_len:
  17. repeat_num = min_clip_len // wave_len + 1
  18. wave_data = wave_data.repeat(1, repeat_num)
  19. wave_data = wave_data[:, :min_clip_len]
  20. else:
  21. clip_num = wave_len // min_clip_len + 1
  22. max_wave_len = clip_num * min_clip_len
  24. wave_data = wave_data.repeat(1, 2)
  25. wave_data = wave_data[:, :max_wave_len]
  27. feats = torch.split(wave_data, min_clip_len, dim = 1)
  28. return feats
import torch
import torchaudio

def load_audio(wave_file):
    wave_data, sr = torchaudio.load(wave_file)
    return wave_data, sr

def save_audio(wave_data, wave_file):
    torch.save(wave_data.clone(), wave_file, _use_new_zipfile_serialization=True)
def get_audio_clips(wave_file, min_sec = 5):
    wave_data, sr = load_audio(wave_file)

    min_clip_len = min_sec * sr

    wave_len = wave_data.shape[1]
    if wave_len < min_clip_len:
        repeat_num = min_clip_len // wave_len + 1
        wave_data = wave_data.repeat(1, repeat_num)
        wave_data = wave_data[:, :min_clip_len]
    else:
        clip_num = wave_len // min_clip_len + 1
        max_wave_len = clip_num * min_clip_len

        wave_data = wave_data.repeat(1, 2)
        wave_data = wave_data[:, :max_wave_len]

    feats = torch.split(wave_data, min_clip_len, dim = 1)
    return feats

We can use this code as follows:

  1. wave_file = "music-jamendo-0039.wav"
  2. feats = get_audio_clips(wave_file)
  3. print(len(feats))
  4. for feat in feats:
  5. print(feat.shape)
wave_file = "music-jamendo-0039.wav"
feats = get_audio_clips(wave_file)
print(len(feats))
for feat in feats:
    print(feat.shape)

In this code, we will split music-jamendo-0039.wav to some 5 seconds small files. Run this code, we will see:

  1. 65
  2. torch.Size([1, 80000])
  3. torch.Size([1, 80000])
  4. torch.Size([1, 80000])
  5. torch.Size([1, 80000])
  6. torch.Size([1, 80000])
  7. torch.Size([1, 80000])
  8. torch.Size([1, 80000])
65
torch.Size([1, 80000])
torch.Size([1, 80000])
torch.Size([1, 80000])
torch.Size([1, 80000])
torch.Size([1, 80000])
torch.Size([1, 80000])
torch.Size([1, 80000])

To understand torch.split() and torchaudio.load(), you can see:

Understand torch.split(): Split Tensor into Chunks – PyTorch Tutorial

Understand torchaudio.load(): Read Audio with Examples – TorchAudio Tutorial