Introduction
SpeechRouter is a unified gateway for speech-to-text: one API in front of 12 providers (Deepgram, Soniox, AssemblyAI, OpenAI, Speechmatics, Azure, AWS, Google, Groq, Mistral, ElevenLabs, Cartesia). You hold one key, learn one schema, and get one bill — and every model on the market is a slug away.
Why it exists
Section titled “Why it exists”Speech vendors disagree on everything: auth schemes, WebSocket dialects, event shapes, timestamp units, pricing units. Teams either marry one vendor or maintain N integrations. SpeechRouter absorbs that variance at the gateway:
- One request shape — the same query params and form fields for every model.
- One event schema — normalized transcripts, words, timestamps, and speakers.
- One failure model — a closed set of error codes, with automatic failover when a provider dies mid-stream.
The three surfaces
Section titled “The three surfaces”| Surface | What it does |
|---|---|
wss://api.speechrouter.ai/v1/listen |
Live streaming transcription |
POST /v1/audio/transcriptions |
Batch file / URL transcription |
GET /v1/models |
The live model catalog with prices and capabilities |
Everything is open source (speech-router/speechrouter,
Apache-2.0) and self-hostable; the hosted gateway at api.speechrouter.ai is the
zero-ops path.
What it costs
Section titled “What it costs”Vendor list price, 0% markup — the catalog carries each vendor’s own pricing in its native unit and we charge exactly that. Details in Pricing & billing.