ovos-stt-http-server Documentation Hub

ovos-stt-http-server exposes any OVOS STT plugin as a network service. On its own it speaks one native protocol. With compat routers enabled it also answers as every major cloud STT API and every popular self-hosted server, so unmodified consumer apps transparently land on this box instead of a third-party cloud.

Goal	Read
Run the server and use the native API	index.md
Make unmodified consumer apps stop calling cloud STT services	voice-pihole.md
Drive this server with the official Python/Node SDK of a specific vendor	api-compatibility.md
Replace an existing self-hosted STT server on the wire	api-compatibility.md
Bridge Home Assistant Voice / Voice PE	wyoming-integration.md
Audio format negotiation	audio-formats.md

Compat coverage

Mounted by create_app() and live on dev (a runnable client script for each is in ../examples/; full reference in api-compatibility.md):

Commercial cloud STT

Vendor	Prefix
OpenAI Whisper	/v1/audio/{transcriptions,translations}
Groq	/groq/openai/v1/audio/transcriptions
Deepgram	/deepgram/v1/listen (HTTP + WS)
Google Cloud STT	/google/v1/speech:recognize
AssemblyAI	/assemblyai/v2/{upload,transcript,realtime/ws}
Gladia	/gladia/v2/{upload,transcription,transcription/{id}}
Speechmatics	/speechmatics/... (REST + WS)
Microsoft Azure Speech	/azure-stt/cognitiveservices/v1 (REST + WS)
AWS Transcribe	/aws/transcribe (batch + streaming WS)
IBM Watson STT	/watson/speech-to-text/v1/recognize (REST + WS)
ElevenLabs Scribe	/elevenlabs/v1/speech-to-text
Wit.ai (Meta)	/wit/speech
Chromium Web Speech	/speech-api/v2/recognize

Self-hosted / open-source server protocols

Server	Surface
whisper.cpp HTTP server	/inference (+ OpenAI alias /v1/audio/transcriptions)
vosk-server (WebRTC)	/vosk-webrtc/offer (needs the aiortc extra)
faster-whisper-server	re-uses /v1/audio/transcriptions
Wyoming (HA)	external adapter, see wyoming-integration.md

OpenAI-compatible Whisper hosts

Apps targeting Groq, Cloudflare Workers AI, Fireworks AI, Together AI, or OpenRouter already speak the OpenAI /v1/audio/transcriptions contract. Point them at this server's /v1 prefix.

Planned / not mounted by default

Present in the package but not mounted by create_app() (need a separate process or CLI flag), tracked for a future release: vosk-server WebSocket (/vosk), vosk gRPC (--vosk-grpc-port), vosk MQTT (--vosk-mqtt-broker), and kaldi-gstreamer-server (/client/ws/speech).

Voice-pihole

voice-pihole.md collects every DNS-rewrite + reverse-proxy

• CA-trust recipe across all the above vendors.

---

TODO / WIP

Streaming partial transcripts

Every compat router with a streaming surface (Deepgram WS, AssemblyAI realtime, Speechmatics RT, Azure WS, Watson WS, vosk WS, AWS streaming, Chromium) runs as buffered batch. Audio is collected for the lifetime of the stream and a single final transcript is emitted on close. OVOS STT plugins don't expose a partial / interim hypothesis interface.

Needed:

• A VAD-driven chunker upstream of the OVOS plugin call, OR
• New plugin-side streaming support (OPM contract change).

Vendor coverage gaps

• Web Speech API native binary protocol: the Sec-WebSocket-Protocol framing modern Chrome/Edge use for the in-browser SpeechRecognition Web API. Distinct from the legacy /speech-api/v2/recognize REST.
• Rev.ai: REST + streaming WS.
• Soniox: gRPC streaming.
• AssemblyAI streaming V3: schema differs from the V2 realtime.
• whisper.cpp verbose_json with segment timestamps: needs VAD pipeline.
• AWS Transcribe streaming over real HTTP/2 event-stream: the WS surface ships a plain WS. SDK-level apps that bypass WS need a translator.
• vosk gRPC StatsService: second proto service not implemented.