lhotse

Why we included this project

Speech researchers often burn more time moving data into a trainable format than they do training. Lhotse handles that layer for you: it turns audio and transcripts into plain-text manifest files, and lets you slice, pad, and augment recordings as flexible units before feeding them to PyTorch dataloaders. It ships ready-made recipes for many common corpora, and its focus has expanded beyond speech to video, image, and text, so a single pipeline can serve multimodal projects. If you are building on the modern Kaldi ecosystem, it works alongside k2 and icefall, but it also integrates with ESPnet and other frameworks.