PPASR

Why we included this project

Chinese teams building speech-to-text have a solid, practical option here. PPASR is an automatic speech recognition framework on PaddlePaddle covering several end-to-end models, DeepSpeech2, Conformer, Squeezeformer, and efficient Conformer variants, and every model can run streaming or non-streaming, so you can pick the latency and accuracy tradeoff that fits live audio or pre-recorded files. The toolkit includes multiple decoders, fbank and mfcc preprocessing, and a long list of data augmentation techniques, so training from scratch is realistic rather than being locked to pretrained weights. A published WenetSpeech pretrained model with character error rates for different decoding setups gives you a concrete baseline before you invest in your own training run. It deploys to servers and Nvidia Jetson devices, and the sentencepiece-based tokenizer leaves the door open for Chinese-English mixed training.