mirror of
https://github.com/hanzoai/speech.git
synced 2026-10-04 07:37:19 +00:00
zeekayandHanzo Dev
6d9aca2b3a
A decoder costs the same for silence as for speech, so an attendant in a quiet room occupied a worker and billed a tenant for text that came back empty. Silero already ships inside faster-whisper and already runs inside every decode; running it in FRONT of the decoder — 0.67 ms per 250 ms push, measured — turns that second into no decode and no charge at all. It is a squelch rather than a per-window classifier, because the classic way to lose more than you save is to start at the trigger: the voice is not detected until it is under way. Sliding a word across a push shows the size of it — with no lead-in, 8 of 32 alignments clip up to 71.8 ms off the front of the word, and with LEAD they keep 428 ms of margin. `duration` on the ack becomes the seconds that reached a decoder, so the plane meters what we listened to. It can exceed one push's own length, on the push that hands over the lead-in, and can never exceed the audio a session was sent. Co-authored-by: Hanzo Dev <[email protected]>
speech
CPU speech for the Hanzo platform: STT (faster-whisper, CTranslate2 int8) and TTS (kokoro-onnx, 82M) behind one OpenAI-shaped surface.
Routes
POST /v1/audio/transcriptions— multipartfile+model(+language,response_format) →{"text": …}POST /v1/audio/speech—{model, input, voice, response_format}→ audio bytes (mp3 default, wav native)GET /v1/models,GET /healthz
Trust model
No auth of its own, in-cluster only. api.hanzo.ai fronts it through the ai
plane, which authenticates and meters every caller — a provider row pointing at
http://speech.hanzo.svc/v1 is the whole integration. Weights bake into the
image at build; the container boots without the network.
Run
uv sync && uv run uvicorn main:app --port 8000
Phase 2 is /v1/realtime: websocket full-duplex (silero-vad segmentation,
partial transcripts, kokoro frames, barge-in) plus camera-frame relay into the
multimodal models.