| # Quickstart |
|
|
| Transcribe a 16 kHz waveform through the streaming CTC path and watch the transcript |
| build up. Memory stays flat over the stream regardless of length. |
|
|
| ```bash |
| pip install -r requirements.txt |
| python run.py --audio sample.wav |
| ``` |
|
|
| ## What this needs |
|
|
| - A 16 kHz mono waveform. If the file is a different rate or stereo, `run.py` resamples |
| and downmixes it. If no file is present, it runs the path on a short silent buffer and |
| prints the expected input shape. |
| - The streaming checkpoint and the SentencePiece BPE (`bpe256.model`, 257 units including |
| the CTC blank). |
|
|
| ## Status |
|
|
| This wires the current shippable streaming checkpoint (460h, 16.4 WER on LibriSpeech |
| test clean). |
|
|
| The streaming loop feeds the audio in chunks and greedy decodes CTC per chunk (no language |
| model). Long streams use a state reset every so many frames (VAD style) to stay in the |
| training regime. Memory stays bounded either way. |
|
|
