k2-training-pipeline

What problem does it solve?

This skill removes the complexity of building end-to-end speech training pipelines by orchestrating data preparation, model training, fine-tuning, and export for ASR and TTS systems so you can produce deployment-ready models with predictable steps and validations.

Core Features & Use Cases

• Data preparation with lhotse: create manifests, cuts, and features for common corpora and custom datasets.
• Training recipes via icefall: run Zipformer, Conformer, and VITS training and fine-tuning with k2 loss (CTC, LF-MMI, pruned RNN-T).
• Model export & deployment: export encoder/decoder/joiner to ONNX or torchscript and validate for sherpa-onnx deployment.
• Use Case: prepare LibriSpeech with lhotse, train a Zipformer transducer with icefall, fine-tune on custom data, then export ONNX models for real-time inference.

Quick Start

Prepare data with lhotse, run icefall training for your chosen recipe, then export the trained model to ONNX and validate it with the recipe's ONNX test script.