atdataset 0.2.0

Audio text dataset for PyTorch training based on webdataset.

ATdataset

A streaming audio-text dataloader for PyTorch training, built on WebDataset.

Designed for large-scale speech recognition training with multi-dataset muxing, dynamic batching, bucketing, and on-the-fly augmentation.

Install

pip install atdataset

For development:

git clone https://github.com/pkufool/ATdataset.git
cd ATdataset
pip install -e ".[dev]"

Quick Start

from atdataset import ATDataloader

dl = ATDataloader(
    datasets="data/tars/train.lst",
    sample_rate=16000,
    max_duration=600.0,
)

for batch in dl:
    audio = batch["audio"]          # (B, T), padded
    audio_lens = batch["audio_lens"]  # (B,)
    feature = batch["feature"]      # (B, T, F), padded
    feature_lens = batch["feature_lens"]  # (B,)
    texts = batch["text"]           # list of str
    break

CLI Tools

After installation, the atdataset command provides two subcommands:

# Build tar shards from a TSV manifest
atdataset build --input train.tsv --output-dir data/tars --num-tars 64

# Generate a .lst file from existing tar shards
atdataset gen_lst --audio-pattern 'data/tars/audios/*.tar' \
    --txt-dir data/tars/manifests --output data/tars/train.lst

Run atdataset build --help or atdataset gen_lst --help for full option details.

Documentation

• Design & Architecture — pipeline architecture, bucketing, muxing, epoch control
• Usage Guide — detailed examples for all features
• Data Format — manifest, tar shard, and lst file formats

License

Apache-2.0