torch-speech-dataloader

A ready-to-use pytorch dataloader for audio classification, speech classification, speaker recognition, etc. with in-GPU augmentations.

• PyTorch speech dataloader with 5 (or less) lines of code. get_torch_speech_dataloader_from_config(config)
• Batch augmentation in GPU, powered by torch-audiomentations
• RIRs augmentation with any set of IR file(s) [cpu]
• MUSAN-like augmentation with any set of source files. Customizable. [cpu]
• Written in one night, may contain bugs!

Install

pip install -U git+https://github.com/zabir-nabil/torch-speech-dataloader.git@main

Use

from torch_speech_dataloader import get_torch_speech_dataloader, get_torch_speech_dataloader_from_config
from torch_speech_dataloader.augmentation_utils import placeholder_gpu_augmentation

config_1 = {
    "filenames" : ["../test.wav"] * 5 + ["../test_hindi.wav"] * 5,
    "speech_labels" : ["test"] * 5 + ["test2"] * 5,
    "batch_size" : 3,
    "num_workers" : 5,
    "device" : torch.device('cuda:1'),
    "sanity_check_path" : "../sanity_test",
    "sanity_check_samples" : 2,
    "batch_audio_augmentation": placeholder_gpu_augmentation,
    "rirs_reverb" : {"apply": True},
    "musan_augmentation" : {"apply": True, "mix_multiples_max_count": -1, "musan_max_len": 1.},
    "verbose" : 0
}

dummy_tsdl = get_torch_speech_dataloader_from_config(config_1)
for d, l in dummy_tsdl.get_batch():
    print(d.shape)
    print(l)

Others

config parameters

• filenames: A list of filepaths for the audio / speech files (usually wav).
• speech_labels: Corresponding labels for filenames / list of audio files.
• batch_size: Batch size of the dataloader.
• num_workers: Dataloader workers.
• device: torch device [default: cpu].
• sanity_check_path: If you want to look at the sample audio files generated, specify a path where the sample augmented audio files will be saved.
• sanity_check_samples: Number of sample audio files to store in the sanity check folder.
• batch_audio_augmentation: Usually, it will run on the GPU batch if gpu device is specified, else on the CPU batch. Any transform (compose) / augmentation, that takes a tensor of dimension [B x C x N].
• rirs_reverb:
  • apply: If apply is true, only then this augmentation will be applied to each audio individually.
  • reverb_source_files_path: A list of IR filepaths.
• musan_augmentation:
  • apply: If apply is true, only then this augmentation will be applied to each audio individually.
  • musan_config: { "music": ([list of music file paths], range_for_num_music_files_to_use, range_for_noise_snr), "speech": ([list of speech file paths], range_for_num_speech_files_to_use, range_for_noise_snr), } [example: augmentation_utils.placeholder_musan_config]
  • mix_multiples_max_count: Multiple noise types should be mixed (music + noise + ...). Number of noise types that should be mixed at most.
  • musan_max_len: <= 0: take the musan noise and crop it with equal length (same as input audio); > 0: maximum length of the cropped musan noise (in secs.).
• audio_augmentation: List of funcs that can be applied to a single audio with shape [N,].
• features: Feature extraction. [N,] -> [T,F].
• feature_augmentation: List of funcs that can be applied to a single feature with shape [T,F].