Icefall

Documentation for icefall, containing speech recognition recipes using k2.

Contents:

• Icefall for dummies tutorial
  • Environment setup
  • Data Preparation
  • Training
  • Decoding
  • Model Export
• Installation
  • (0) Install CUDA toolkit and cuDNN
  • (1) Install torch and torchaudio
  • (2) Install k2
  • (3) Install lhotse
  • (4) Download icefall
  • Installation example
  • Test Your Installation
  • YouTube Video
• Docker
  • Introduction
  • View available tags
  • Download a docker image (CUDA)
  • Download a docker image (CPU)
  • Run a docker image with GPU
  • Run a docker image with CPU
  • Run yesno within a docker container
• Frequently Asked Questions (FAQs)
  • OSError: libtorch_hip.so: cannot open shared object file: no such file or directory
  • AttributeError: module ‘distutils’ has no attribute ‘version’
  • ImportError: libpython3.10.so.1.0: cannot open shared object file: No such file or directory
• Model export
  • Export model.state_dict()
  • Export model with torch.jit.trace()
  • Export model with torch.jit.script()
  • Export to ONNX
  • Export to ncnn
• FST-based forced alignment
  • Two approaches
  • Kaldi-based forced alignment
  • k2-based forced alignment

• Recipes
  • Non Streaming ASR
    • aishell
    • LibriSpeech
    • TIMIT
    • YesNo
  • Streaming ASR
    • Introduction
    • LibriSpeech
  • RNN-LM
    • Train an RNN language model
  • TTS
    • VITS-LJSpeech
    • VITS-VCTK
  • Fine-tune a pre-trained model
    • Finetune from a supervised pre-trained Zipformer model
    • Finetune from a pre-trained Zipformer model with adapters

• Contributing
  • Contributing to Documentation
  • Follow the code style
  • How to create a recipe
• Huggingface
  • Pre-trained models
  • Huggingface spaces

• Decoding with language models
  • Shallow fusion for Transducer
  • LODR for RNN Transducer
  • LM rescoring for Transducer