| model_type: "NemoSTT" | |
| # frame size in ms for incremental transcription | |
| frame_size: 1000 | |
| # Parameters from https://github.com/NVIDIA/NeMo/blob/stable/tutorials/asr/Online_ASR_Microphone_Demo.ipynb | |
| frame_overlap: 2 | |
| offset: 4 | |
| # timestep_duration = model._cfg.preprocessor['window_stride'] | |
| # for block in model._cfg.encoder['jasper']: | |
| # timestep_duration *= block['stride'][0] ** block['repeat'] | |
| timestep_duration: 0.02 | |
| # Sample rate | |
| sample_rate: 16000 | |
| # Minimum detectable VAD section in ms | |
| min_speech_duration: 400 | |
| # Timeout in ms to flush results if speech wasn't finished semantically | |
| max_silence_duration: 1000 | |
| # VAD frame size in ms | |
| vad_frame_ms: 20 | |
| transcribe_realtime: False | |
| predict_punctuation: False | |
| alpha: 0.0253813572180912 | |
| beta: 0.08 | |