Server uses the sherpa_onnx project to run models on the zipformer architecture. Available models can be found here.
I'm still looking for a better implementation of AGC for loudness normalization, as various diy satellites can have low levels. But in some situations, it may not work very well. Use a handler that saves an audio sample to evaluate the results.
Command dictionary [optional] - if specified, after recognizing command-max-words words, a dictionary check is performed, and if a match is found, it will become the final recognition result. The function works in conjunction with this custom client (it is capable of stopping audio transmission on command from the server).
The model directory must contain 4 files: encoder.onnx decoder.onnx joiner.onnx tokens.txt
Required libraries: pip install sherpa-onnx wyoming numpy
Example launch for Windows:
python -m wyoming_streaming_asr --model-dir "\models\vosk-ru" --language ru --command-file commands.txt --command-max-words 5
Latest Russian models from Alpha Cephei https://huggingface.co/alphacep/
Demo with ru model.
The streaming version of Parakeet (en) has been released. Run this model with the --decoding-method greedy_search option. The model was converted to onnx with the 1.12s setting.