Skip to content
Source
Sign inSign up
TanStack/ai

feat: live speech-to-text sessions for voice and telephony agents #1944

Send to my agent
Opentombeckenham opened this issue 0 comments
tombeckenhamcommented

Problem

TanStack AI can't turn live audio into text on the server. That blocks the whole voice and telephony agent space: phone agents, call centres, IVR replacement, meeting bots, and voice agents behind an SFU.

  • generateTranscription() only takes a finished file.
  • Realtime voice (useRealtimeChat) connects the browser straight to the provider. The connection reads the microphone itself (startAudioCapture), so a server can't feed it audio from a phone call.

A server-owned voice agent needs to feed audio as it arrives, and get text back while the caller talks. The audio comes from Twilio Media Streams, Telnyx, Plivo, Vonage, LiveKit, an SFU, or a browser WebSocket. Today teams stitch this together from vendor SDKs, then hand the text to chat().

This came up from a user who runs a voice agent on Cloudflare agents with withVoice. Their reply text already comes from chat(). Speech-to-text is the piece TanStack can't provide. See Transcriber in agents/voice for one pipeline's view of the shape.

Proposal

A new activity, a live transcription session, with one provider-neutral adapter contract.

import { transcriptionSession } from '@tanstack/ai'
import { deepgramLiveTranscription } from '@tanstack/ai-deepgram'

const session = await transcriptionSession({
  adapter: deepgramLiveTranscription({ model: 'nova-3' }),
  audio: { encoding: 'mulaw', sampleRate: 8000, channels: 1 }, // a Twilio call
  language: 'en',
})

twilioSocket.on('media', (m) => session.send(m.payload)) // raw audio chunks

for await (const event of session) {
  // { type: 'interim', text }  text that may still change
  // { type: 'final', text }    stable text for a segment
  // { type: 'turn-end' }       the provider thinks the caller finished talking
  // { type: 'error', error }
}

await session.close()

Swap the adapter to change provider. Nothing else changes.

Goals

  • Runs on the server: Node, Bun, Workers and Durable Objects. No browser APIs.
  • One small adapter contract (open, send audio, receive events, close), so providers are cheap to add, including community adapters.
  • Telephony audio is first-class: G.711 μ-law and A-law at 8 kHz, plus linear PCM at 8, 16, 24 and 48 kHz. Each adapter declares what it accepts. A shared helper converts formats when the provider can't take the input as is.
  • Optional turn-end signal, when the provider has endpointing.
  • Barge-in support: the caller's speech during agent playback must reach the app fast, as interim events.
  • Small enough to wrap for other voice pipelines, for example an agents/voice Transcriber, in a few lines.

Providers

The first PR ships at least two providers from different vendors. That proves the contract isn't shaped around one API.

ProviderPackageNotes
Deepgram (Nova-3, Flux)new @tanstack/ai-deepgramMost common STT in telephony stacks. Flux has turn detection
AssemblyAI (Universal-Streaming)new @tanstack/ai-assemblyaiTurn detection built in
ElevenLabs Scribe realtime@tanstack/ai-elevenlabsPackage exists
OpenAI realtime transcription@tanstack/ai-openaiPackage exists
Workers AI (Deepgram models over the AI binding)@tanstack/ai-cloudflarePackage exists. The user who asked
Gladia, Speechmatics, Soniox, Google, Azurecommunity or later

I haven't confirmed live-streaming support or exact model ids for each row. Check before you build.

Out of scope for the first PR

  • A browser hook (useLiveTranscription). Add it later if people ask.
  • Streaming text-to-speech audio. Telephony needs it too (μ-law out, sent as it's made), but it's a separate issue. generateSpeech() returns whole clips today.
  • Carrier helpers (Twilio, Telnyx and similar message parsing). Document the pattern first.
  • A full server voice loop (turn-taking, barge-in, interruption). This issue is the building block for that.

Open questions

  1. An async iterator of events, or on(event, handler) callbacks like RealtimeConnection?
  2. Should events follow AG-UI (CUSTOM events), or stay a plain transcription event union?
  3. Where does format conversion live: in @tanstack/ai, or in a small @tanstack/ai-utils audio helper?
  4. E2E: aimock has no live-audio WebSocket mocking. Use unit tests with a fake adapter until it does?

Done when

  • Activity and adapter contract in @tanstack/ai, plus two or more providers from different vendors.
  • A docs page under docs/media/, with a telephony example (a Twilio Media Streams call into chat()).
  • A media-generation skill update.
  • Changesets.
  1. tombeckenham changed the title

  2. github-actions[bot](GitHub profile, opens in a new tab) added labelswaiting-on: maintainer

  3. tombeckenham unassignedAlemTuzlak

  4. tombeckenham assignedtombeckenham

  5. tombeckenham linked this to #1964

Sign in to join this conversation.