Add Voice

Add Grok Speech to Speech to an existing app. Run on /add-voice, typed Voice Mode, or clear “add Grok voice” intent.

Goal

Working duplex path: user-app mic in, audio out, wss://api.x.ai/v1/realtime?model=grok-voice-latest, safe auth. Cursor has no native mic; wire the app (or a sample client), not the IDE.

Protocol first

Language-agnostic event loop. TypeScript samples default. Short Python twins only where the client API differs (e.g. ws vs websockets).

Docs

https://docs.x.ai/developers/model-capabilities/audio/speech-to-speech
https://docs.x.ai/developers/model-capabilities/audio/ephemeral-tokens
Pricing (cite docs only): https://docs.x.ai/developers/pricing (~$0.08/min STS + $0.004/text item; max session 120 min)

Steps

Map the app
- Stack: none, OpenAI Realtime, STT→LLM→TTS cascade, TTS/STT only.
- Client: web / Node / iOS / Android / server.
- If a cascade or OpenAI Realtime exists: replace it with the single duplex loop below (URL, model, voice, event diffs); keep standalone /v1/stt or /v1/tts only if the product still needs one-shot listen or speak outside the agent.
Auth
- Server: Bearer XAI_API_KEY.
- Browser/mobile: backend POST https://api.x.ai/v1/realtime/client_secrets, client uses ephemeral token (Bearer or browser sec-websocket-protocol: xai-client-secret.<token>).
- Never put a long-lived key in client bundles. Do not paste keys in chat.
Connect + session
- URL: wss://api.x.ai/v1/realtime?model=grok-voice-latest
- On open: session.update with voice (default eve), instructions, turn_detection: { type: "server_vad" } (or null for push-to-talk), PCM 24 kHz unless the app already standardizes elsewhere.
- Set audio.input.transcription.model: "grok-transcribe" or no user transcript arrives (conversation.item.input_audio_transcription.updated is cumulative, not a delta).
- Tools if needed: web_search, x_search, file_search, mcp, custom function.
Audio I/O (app-side)
- One AudioContext per session for capture and playback, created inside the user gesture (autoplay policy). Ask for 24 kHz; if the browser gives another rate, resample before sending.
- Mic → AudioWorklet in ~100 ms chunks → input_audio_buffer.append (or binary transport). Start WS and mic in parallel; buffer early audio, flush on open.
- Play response.output_audio.delta immediately; schedule with a ~150 ms lead so chunks butt together. On input_audio_buffer.speech_started, stop everything queued (barge-in).
- Transcript rows: create the user row on input_audio_buffer.committed (item_id), fill it on …transcription.updated; assistant text from response.output_audio_transcript.delta / .done, close the turn on response.done.
- On function tools: function_call_output, finish playback, then response.create.
Composer UI convention
- One primary button, right side of the composer. Empty composer → waveform icon (stroked, e.g. Phosphor WaveformIcon weight="bold"; never the fill weight, which renders as a blob at 16 px), starts voice mode. Any text present → classic send arrow; in voice mode that text goes into the live session (conversation.item.create + response.create). Text reply streaming → stop square.
- While voice is live the same button shows an animated waveform (4 bars, ~3 px wide, 2 px gap, ~16 px tall, min scale 0.4 so they stay legible in a 28 px button) and ends the session on click. Phase drives the animation: listening slow, speaking fast, connecting/thinking slower and slightly dimmed (opacity ≥ 0.75). Honor prefers-reduced-motion. No X button, no pulsing ring.
- Status lives in the composer, not around it: the placeholder reads Connecting… / Listening… / Thinking… / Speaking…, plus an sr-only role="status". No separate status row.
- The microphone icon is reserved for dictation (/add-dictation). Never use it for voice mode.
TS skeleton (default)

const url = "wss://api.x.ai/v1/realtime?model=grok-voice-latest";
// Node: pass Authorization header. Browser: use xai-client-secret.<token> protocol.
const ws = new WebSocket(url /* , { headers: { Authorization: `Bearer ${token}` } } */);

ws.addEventListener("open", () => {
  ws.send(JSON.stringify({
    type: "session.update",
    session: {
      voice: "eve",
      instructions: "You are a helpful voice agent.",
      turn_detection: { type: "server_vad" },
    },
  }));
});

ws.addEventListener("message", (ev) => {
  const event = JSON.parse(String(ev.data));
  if (event.type === "response.output_audio.delta") {
    // decode base64 PCM and play
  }
});

Python twin (only if the app is Python)

import json, os, websockets

url = "wss://api.x.ai/v1/realtime?model=grok-voice-latest"
headers = {"Authorization": f"Bearer {os.environ['XAI_API_KEY']}"}

async with websockets.connect(url, additional_headers=headers) as ws:
    await ws.send(json.dumps({
        "type": "session.update",
        "session": {
            "voice": "eve",
            "instructions": "You are a helpful voice agent.",
            "turn_detection": {"type": "server_vad"},
        },
    }))
    async for raw in ws:
        event = json.loads(raw)
        if event.get("type") == "response.output_audio.delta":
            pass  # decode and play

Instrument (before the first human test)
- Run /debug-voice: it proposes a plan, then installs a dev-only log sink (POST /api/voice/log → .voice-logs/<sessionId>.ndjson, gitignored), a client logger with audio reduced to byte counts, and the session id in the UI, in the app's own language.
- The same skill carries the fix loop and the symptom → log signature → fix table.
Smoke
- Text turn via conversation.item.create + response.create; confirm audio or transcript events.
- Confirm no long-lived key in client (grep the built client bundle for the env name and client_secrets).
- Hand the app to the user with headphones. On speakers the mic hears the reply and the model answers itself; that is echo, not a bug in the loop.
- Iterate with /debug-voice.

Out of scope

Speech-to-text only (/add-dictation), speaking text (/add-read-aloud)
Image generation and text-only inference
Invented endpoints, events, or CLI flags