PPASR is an end-to-end Chinese automatic speech recognition (ASR) framework built on PaddlePaddle. It offers a simplified and practical approach to ASR, supporting popular models like DeepSpeech2, Conformer, and Squeezeformer, with both streaming and non-streaming capabilities. The project targets developers and researchers looking for a flexible and efficient ASR solution that can be deployed on servers and edge devices like Nvidia Jetson.

How It Works

PPASR (V3) is a significant overhaul from V2, focusing on ease of use and enhanced performance. It utilizes kaldi_native_fbank for faster, multi-platform compatible audio preprocessing and sentencepiece for tokenization, enabling easier handling of multiple languages and mixed-language training. The framework supports various decoding methods (e.g., ctc_greedy_search, ctc_prefix_beam_search, attention_rescoring) and data augmentation techniques for improved robustness.

Quick Start & Requirements

• Installation: Requires Anaconda 3, Python 3.11, and PaddlePaddle 2.6.1.
• OS: Windows 11 or Ubuntu 22.04.
• Resources: Pre-trained models are available via a "knowledge planet" (details in README).
• Docs: Links to online demos, WeChat mini-programs, and video tutorials are provided in the README.

Highlighted Details

• Supports multiple ASR models including Conformer, Squeezeformer, and DeepSpeech2, all with streaming and non-streaming options.
• Offers diverse decoding strategies and data augmentation methods for customized performance.
• Pre-trained models are available for various datasets (WenetSpeech, AIShell, Librispeech) with reported character error rates (CER) or word error rates (WER). For example, a Conformer model with attention_rescoring on WenetSpeech achieves a CER of 0.13786 on test_net.

Maintenance & Community

• The project is actively developed, with V3 being the latest release.
• Community discussion is encouraged via QQ groups and a "knowledge planet" which also provides model files. Links to Bilibili for video explanations are included.

Licensing & Compatibility

• The README does not explicitly state a license. Compatibility for commercial use is not specified.

Limitations & Caveats

• Pre-trained model weights are not directly downloadable from the repository and require joining a paid "knowledge planet."
• The project is primarily focused on Chinese ASR, though it mentions support for English and mixed-language training.