Deepgram compatibility

The gateway natively speaks the Deepgram dialect — live and prerecorded — on the same /v1/listen endpoints. Existing Deepgram code migrates by changing the URL and the key, nothing else.

How dialect selection works

The auth scheme picks the dialect:

• Authorization: Token <key> → Deepgram dialect (what the Deepgram SDK sends)
• Authorization: Bearer <key> → native SpeechRouter schema

Live streaming with @deepgram/sdk

import { createClient, LiveTranscriptionEvents } from '@deepgram/sdk'

const dg = createClient(SPEECHROUTER_KEY, {
  global: { websocket: { options: { url: 'wss://api.speechrouter.ai' } } },
})

const live = dg.listen.live({ model: 'nova-3', interim_results: true })
live.on(LiveTranscriptionEvents.Transcript, (t) => {
  console.log(t.channel.alternatives[0].transcript)
})

You get Deepgram-shaped events — Results, SpeechStarted, UtteranceEnd, Metadata — with is_final / speech_final semantics preserved, plus CloseStream and KeepAlive handled as Deepgram defines them.

The unlock: model accepts any SpeechRouter slug. model: 'nova-3' routes to Deepgram as usual; model: 'soniox/stt-rt-v5' routes to Soniox — through the unchanged Deepgram SDK.

Prerecorded

POST /v1/listen with Token auth accepts Deepgram’s prerecorded shape and returns Deepgram-shaped JSON (results.channels[0].alternatives[0]).

Working example

A complete Next.js app using the Deepgram SDK against SpeechRouter, with a model picker across providers: speech-router/deepgram-sdk-example.