DOCSEARCHbyAlgolia
Submit Docs

icefall

repository·master·Indexed 23 days ago

https://github.com/k2-fsa/icefall

A repository of speech recognition recipes that leverages k2 and lhotse for training, serving as a bridge between dataset-specific training procedures and deployment frameworks like sherpa. It includes recipes for ASR models such as Conformer CTC, TDNN-LSTM-CTC, and various RNN-Transducer variants (including stateless and modified stateless) for datasets like Aishell, Aishell2, and Aidatatang_200zh.

Tokens
175.7K
Snippets
478
Records
655
Agent score
80%

What's inside icefall

  1. Overview of the tiny_transducer_ctc recipe

    master

    The tiny_transducer_ctc recipe is designed for streaming Automatic Speech Recognition (ASR) on low-cost devices. It features models with 1-2M parameters using a small convolutional encoder. The training process combines Transducer and CTC losses and supports both phone and BPE lexicons.

    Key Architectural Features:

    • Encoder: Consists of 2 subsampling layers followed by a stack of Conv1d-batchnorm-activation-causal_squeeze_excite blocks.
    • Latency Optimization: To reduce latency, half of the blocks use causal convolution (at the cost of slightly higher WER).
    • Squeeze-and-Excitation: Uses a moving average filter instead of global average pooling to maintain causality for streaming.
    • Lexicon Usage:
      • Phone lexicon: Performs better for CTC decoding (with HLG) but worse for transducer decoding.
      • BPE lexicon: Generally preferred for transducer decoding.
    • SpecAugment: Not recommended for very small models as they tend to underfit. For larger models, a less aggressive SpecAugment is suggested.

  2. Overview of the Yesno ASR recipe

    master
    The yesno recipe is the simplest Automatic Speech Recognition (ASR) recipe available in icefall. It is designed for minimal complexity and can be executed on a CPU, typically completing in under 30 seconds. This recipe is suitable for testing the icefall environment or understanding the basic ASR pipeline structure.

  3. Overview of Aishell2 ASR recipes

    master
    The aishell2 recipe provides various Automatic Speech Recognition (ASR) models trained on the AISHELL-2 dataset (1000 hours of clean read-speech). This recipe is designed for industrial-scale applications and includes features such as Chinese word segmentation, flexible vocabulary expansion, and phone set transformation. It supports state-of-the-art techniques like time-delayed neural networks and Lattice-Free MMI (LF-MMI).

  4. Overview of HENT-SRT Speech-to-Text Translation Recipe

    master
    HENT-SRT is a speech-to-text translation (ST) recipe based on the paper HENT-SRT: Hierarchical Efficient Neural Transducer with Self-Distillation for Joint Speech Recognition and Translation. It implements a system using a Zipformer encoder with a pruned transducer and a stateless decoder. The recipe is designed for multi-way parallel ST tasks using conversational datasets.

  8. Overview of the XBMU-AMDO31 ASR recipe

    master
    The xbmu_amdo31 recipe provides Automatic Speech Recognition (ASR) models trained on the XBMU-AMDO31 corpus. This corpus is an open-source Amdo Tibetan speech dataset published by Northwest Minzu University, consisting of 31 hours of speech data. The dataset includes transcribed texts and a Tibetan pronunciation lexicon (based on the Lhasa dialect) to support training Amdo Tibetan ASR systems.

  10. Overview of the Fluent Speech Commands recipe

    master

    The Fluent Speech Commands recipe is designed for a dataset that transcribes short speech utterances into structured action frames. For example, an utterance like "turn the lights on in the kitchen" is transcribed into a frame such as {"action": "activate", "object": "lights", "location": "kitchen"}.

    Dataset Statistics:

    • Training set: 23,132 utterances
    • Test set: 3,793 utterances

    For more details on the dataset, refer to the Dataset Paper.

  11. Overview of ReazonSpeech ASR recipes

    master

    The egs/reazonspeech/ASR/ directory contains Automatic Speech Recognition (ASR) recipes for the ReazonSpeech dataset, which consists of over 35,000 hours of natural Japanese speech from terrestrial television.

    This directory specifically provides implementations for Transducers. The available transducer variants differ by their encoder and decoder architectures:

  12. Overview of the multi_ja_en bilingual ASR model

    master
    The multi_ja_en recipe provides a bilingual Japanese-English Automatic Speech Recognition (ASR) model. This model is built using ReazonSpeech, an open-source dataset containing over 35,000 hours of natural Japanese speech collected from terrestrial television streams. The model is designed to handle both Japanese and English speech by training on a combination of ReazonSpeech and LibriSpeech datasets.