JavaScript / TypeScript

npm install speechrouter

Works in Node ≥ 18 and modern browsers (native fetch + WebSocket). Fully typed, zero dependencies.

Client

import { SpeechRouter } from 'speechrouter'

const sr = new SpeechRouter({ apiKey: process.env.SPEECHROUTER_API_KEY! })
// self-hosted? new SpeechRouter({ apiKey, baseUrl: 'https://gateway.internal' })

Batch

const { text } = await sr.transcribe({
  model: 'openai/whisper-1',
  file: audioBlob,                       // File | Blob | ArrayBuffer | Uint8Array
})

// typed by responseFormat:
const verbose = await sr.transcribe({
  model: 'assemblyai/universal-2',
  file: audioBlob,
  responseFormat: 'verbose_json',        // → words[], duration, language
  diarization: true,
})

Streaming

const stream = sr.listen({
  model: 'soniox/stt-rt-v5',
  fallbacks: ['deepgram/nova-3'],
  sampleRate: 16000,
})

stream.on('transcript', (t) => t.is_final && render(t.text))
stream.on('utterance_end', () => agent.respond())
stream.on('provider_switched', (e) => log(`failover → ${e.to}`))

stream.sendAudio(pcm)          // Uint8Array / ArrayBuffer of 16-bit PCM
await stream.finalize()        // flush finals, resolve on done

Events are also an async iterable:

for await (const event of stream) {
  if (event.type === 'transcript' && event.is_final) …
}

The client keepalives idle sockets automatically.

Browser: client tokens

// server route
const { token } = await sr.createToken({ ttlSeconds: 60 })

// browser — token instead of the key, via subprotocol auth
const stream = new SpeechRouter({ apiKey: token }).listen({ model: 'deepgram/nova-3' })

Errors

Failures throw SpeechRouterError with the gateway’s stable code on .code.