lhotse-speech/lhotse

Health v2 · maintenance only

Adoption not part of the score

1149 stars · 278 forks observed · 2026-09-01

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Lhotse is a Python library for flexible, scalable preparation of multimodal (speech, audio, video, image, text) data for machine learning, part of the next-generation Kaldi ecosystem. It provides data preparation recipes, efficient dataloading with bucketing and dataset blending, and sequential I/O formats for distributed training.

Use cases

• prepare speech corpora for ASR model training
• build multimodal audio-text training pipelines in PyTorch
• blend and bucket audio datasets for efficient dataloading
• shard audio data for distributed multi-node training
• convert Kaldi data directories to a Python-friendly format
• deduplicate and randomize training data across GPUs

When to choose

• you train speech or audio models in PyTorch and need scalable data pipelines
• you want Python-centric alternatives to Kaldi data preparation
• you need efficient on-the-fly bucketing or dataset blending for large audio corpora

When to avoid

• you need a ready-made speech recognition model rather than data tooling
• your project is not Python/PyTorch based
• you only need simple one-off audio file conversions

Facets

library · maturity active

machine-learning audio-processing etl data-science sdk speech-processing machine-learning python cross-platform speech-recognition kaldi pytorch dataloader multimodal data-preparation audio-cuts dataset-blending audio data-engineering natural-language-processing

Member repositories

For agents

markdown · JSON · MCP: product_card(name="lhotse-speech/lhotse")