Introduction

Umay speech recognition: files, live audio, voice features

Umay is Mihu’s speech recognition service: files and live audio, Turkish first. Audio and transcripts stay on the server of the region you call.

GET /health lists the languages and models a server has.

Regions

Region	Base URL	WebSocket
Europe	https://stt.eu.mihu.ai	wss://stt.eu.mihu.ai
Türkiye	https://stt.tr.mihu.ai	wss://stt.tr.mihu.ai

Choose the region when you call the API: the two regions have the same paths, fields and answers, and your data stays on the server of the region you call. Examples use the Europe address. Every REST endpoint has a page under API Reference where you can pick the region, send a request and see the answer.

Quick start

Europe

curl https://stt.eu.mihu.ai/v1/audio/transcriptions -H "Authorization: Bearer $KEY" \
  -F file=@call.wav -F model=umay -F language=tr

{"text": "Merhaba siparişiniz yarın saat 14'te teslim edilecek.", "model": "umay", "language": "tr", "duration": 4.36, "processing_time": 0.124}

Two APIs on the same server, side by side:

• Native API — model=umay / umay-stream (or no model): see Native API.
• OpenAI-compatible API — the official OpenAI SDKs work by changing base_url. See OpenAI-compatible API.

The model name in the request decides which one answers; all Umay options (normalize, gender_detection, diarize, hints, …) work with both.

Authentication

Every request needs the API key: header Authorization: Bearer <KEY>, or in the URL ?api_key=<KEY> (for WebSocket clients that cannot set headers, e.g. browsers). Browser OpenAI Realtime clients may also send it as the WebSocket subprotocol openai-insecure-api-key.<KEY> (next to realtime). Missing or wrong key → 401 (native body {"error":"invalid api key"}; OpenAI’s error object on /v1/models, /v1/realtime and for OpenAI SDK clients).

Endpoints

Method	Path	Use
GET	/health	Languages and loaded models
POST	/v1/audio/transcriptions	Transcribe a file (multipart) — native or OpenAI-compatible by model; with webhook_url: asynchronous
GET	/v1/audio/transcriptions/{id}	Status and result of an asynchronous transcription
WS	/v1/stream	Live transcription, native protocol (voice agents, calls)
WS	/v1/realtime	Live transcription, OpenAI Realtime transcription session
GET	/v1/models	Model list (OpenAI list shape)
POST	/v1/voiceprints	Store one person’s voiceprint (with their consent) for reference_voiceprint
GET	/v1/voiceprints, /v1/voiceprints/{id}	List your voiceprints / one voiceprint
DELETE	/v1/voiceprints/{id}	Delete a voiceprint for good
POST	/v1/voiceprints/verify	Speaker verification: is this recording the voiceprint’s person? Signed result
GET	/v1/voiceprints/jwks	Public key of the verification signatures
POST	/v1/audio/isolation	Audio isolation: the recording back without background voices and noise
GET, DELETE	/v1/audio/isolation/{id}, /v1/audio/isolation/{id}/audio	Status, download and deletion of an asynchronous isolation job