datasets
Everything about datasets in the SpeeChain toolkit.
This sub-package is the single place for all dataset-related code:
abs.py: the abstractDatasetbase class that reads data instances from the disk into memory and packages them into batches.speech_text.py: the built-inSpeechTextDatasetimplementation used by speech-text tasks (ASR, TTS, etc.).meta_generator.py: the abstract base class for per-dataset metadata generation scripts (data/{dataset_name}/meta_generator.py).meta_post_processor.py: the abstract base class for per-dataset metadata post-processing scripts (data/{dataset_name}/meta_post_processor.py).pyscripts/: the fixed executable scripts used bydata/data_dumping.shanddata/mfa_preparation.sh(feature extraction, waveform downsampling, vocabulary generation, etc.).
The dataset-dumping code lives inside the speechain package (instead of
the top-level data folder) so that:
- it does not shadow the HuggingFace
datasetspackage, and - it is covered by the CI checks (Black & Ruff) that only scan the
speechaindirectory.