Compare model

Live voice

OpenAI logoGPT Live 1 Transcribe

TokenRex's speech-to-text mode built on GPT Live 1: streaming microphone transcripts with client-muted playback.

gpt-live-1-transcribe

Best for: Real-time voice agents and phone-style calls

Share

Catalog status

Enabled

Research checked

2026-10-30

Benchmark results

No recorded results

Results are source-attributed, not TokenRex measurements. Different tests and thinking variants are not interchangeable.

Published benchmark coverage

Reviewed sources do not include results for this exact version. Provider specifications and API capabilities remain listed separately.

Model specifications

Underlying model

GPT Live 1

Documented capability

Mode developed by

TokenRex

Documented capability

Input and output modalities

Provider-reported formats; API compatibility is detailed below.

Input

Live microphone audio

Output

Streaming text transcripts
Exact model ID
gpt-live-1-transcribe
Model type
Live voice
Underlying model
GPT Live 1
Mode developed by
TokenRex
Task
Speech-to-text (STT)
Transport
Live WebSocket signaling + WebRTC audio

Exact token limits are not published in the reviewed sources, so no estimated limits are shown.

Live model mode

One enabled version is currently listed for this model.

GPT Live 1 Transcribe — Speech-to-text (STT)

TokenRex's live speech-to-text mode built on GPT Live 1. Microphone audio becomes streaming transcripts; playback must be muted by the client.

Voice and Transcribe are modified modes made by TokenRex, not separately trained OpenAI models or native standalone TTS/STT endpoints. The underlying GPT Live 1 model still runs. Only Voice carries the Beta label. Reading may skip or change words, and transcription may produce attempted responses; exact reading, permanent silence and instant speech are not guaranteed.

What it is used for

Live captions, dictation, meeting notes and voice-input interfaces.

How Speech-to-text (STT) works

  1. Use a TokenRex API key privately through your server and choose
    gpt-live-1-transcribe
    .
  2. Connect to wss://tokenrex.ai/api/public/v1/live?model=gpt-live-1-transcribe, then send session.start with a WebRTC offer. Keep gateway.heartbeat running every second.
  3. Send microphone audio and read session.input_transcript.delta events. Never attach incoming audio to a speaker. Surface any attempted model output as a warning.
  4. Send session.close when finished and retain the final usage event. Idle time is billable; never reconnect or replay automatically.
// Run only after the user explicitly clicks Start.
// Your server must proxy this connection and keep TOKENREX_API_KEY private.
// Upstream: wss://tokenrex.ai/api/public/v1/live?model=gpt-live-1-transcribe
const ws = new WebSocket("wss://YOUR_APP_HOST/live?model=gpt-live-1-transcribe");
const pc = new RTCPeerConnection();
pc.createDataChannel("oai-events");
const mic = await navigator.mediaDevices.getUserMedia({ audio: true });
mic.getTracks().forEach(track => pc.addTrack(track, mic));
// Do not attach incoming audio to a speaker: this mode is transcription-only.
pc.ontrack = () => {};
await pc.setLocalDescription(await pc.createOffer());
await new Promise(resolve => pc.iceGatheringState === "complete" ? resolve() :
  pc.addEventListener("icegatheringstatechange", () => {
    if (pc.iceGatheringState === "complete") resolve();
  }));
await new Promise((resolve, reject) => {
  if (ws.readyState === WebSocket.OPEN) resolve();
  else { ws.onopen = resolve; ws.onerror = reject; }
});
let heartbeat;
ws.onmessage = async ({ data }) => {
  const event = JSON.parse(data);
  if (event.type === "gateway.session.created") {
    await pc.setRemoteDescription({ type: "answer", sdp: event.transport.sdp });
    heartbeat = setInterval(() => ws.send(JSON.stringify({ type: "gateway.heartbeat" })), 1000);
  }
  if (event.type === "session.input_transcript.delta") console.log(event.delta);
  // Expose attempted output as a warning; instructions cannot guarantee silence.
  if (event.type === "session.output_transcript.delta") console.warn("Unexpected output", event.delta);
  if (event.type === "session.usage.updated") console.log("Billed seconds", event.usage.seconds);
  if (event.type === "tokenrex.error") console.error(event.error.message);
};
ws.send(JSON.stringify({
  type: "session.start",
  session: { audio: { output: { voice: "marin" } } },
  tools: [], transport: { type: "webrtc", sdp: pc.localDescription.sdp },
}));

function stop() {
  clearInterval(heartbeat);
  ws.send(JSON.stringify({ type: "session.close" }));
  mic.getTracks().forEach(track => track.stop());
  pc.close();
  // Keep the socket until the final usage event; never reconnect automatically.
}

Pricing — billed per second, not per character

Current TokenRex rate: $0.00058 per second, equivalent to $0.035 per minute. The minimum is 15 seconds per session ($0.00875). A minute is an equivalent, not a billing block. All connected time, including silence and waiting between readings, is billable. Muting playback does not lower the rate. Any separately requested reasoning or tool usage has its own charges.

Session durationEstimated charge
15-second minimum$0.00875
20 seconds$0.0117
1 minute$0.035
10 minutes$0.35
1 hour$2.10

Voices, languages and accents

Discover sourced voice IDs through GET /api/public/v1/models → this model → live.voices. Use session.audio.output.voice (default marin); voice selection is fixed for the call. Languages and accents are instruction preferences, not native speed controls. The first 22 voices were session-tested; additional documented voices have not been individually auditioned. Transcribe does not need a spoken output voice.

VoiceAccentCharacter
Marin (marin)Neutral (default)Feminine
Cedar (cedar)NeutralMasculine
Alloy (alloy)NeutralNeutral
Ash (ash)NeutralMasculine
Ballad (ballad)NeutralMasculine
Coral (coral)NeutralFeminine
Echo (echo)NeutralMasculine
Sage (sage)NeutralFeminine
Shimmer (shimmer)NeutralFeminine
Verse (verse)NeutralMasculine
Quartz (quartz)English / AustralianFeminine
Ripple (ripple)English / AustralianMasculine
Vesper (vesper)English / BritishMasculine
Willow (willow)English / IrishFeminine
Stone (stone)English / IrishMasculine
Gleam (gleam)English / North AmericanFeminine
Meridian (meridian)English / North AmericanMasculine
Bossa (bossa)Portuguese / BrazilianFeminine
Tempo (tempo)Portuguese / BrazilianMasculine
Beacon (beacon)English / FilipinoMasculine
Delta (delta)English / Southern USFeminine
Cinder (cinder)English / Southern USMasculine
Nuri (nuri)GermanFeminine
Juni (juni)GermanFeminine
Flitz (flitz)GermanMasculine
Brise (brise)FrenchFeminine
Sillage (sillage)FrenchMasculine
Nira (nira)HindiFeminine
Harema (harema)JapaneseFeminine
Noeul (noeul)KoreanFeminine

Features and tool support

TokenRex’s exposed interface, not every feature advertised by the provider. “Available” describes the implemented interface, not a successful test of every input or tool.

Input through TokenRex

Available

Live microphone audio over WebRTC; session start and control events over wss://tokenrex.ai/api/public/v1/live. Tested: sessions opened for all 22 voices.

Output through TokenRex

Available

Spoken audio streamed back in real time, plus live transcript events for both sides of the call.

Function tools

Available

Send your own tools with session.start; TokenRex forwards each call to your app and returns your verified result to the conversation.

Languages

Available

Any voice can speak Hindi, Hinglish, English and other languages — set the rule in session.instructions.

Underlying model

Limited

GPT Live 1

Mode developed by

Limited

TokenRex

Task

Limited

Speech-to-text (STT)

Transport

Limited

Live WebSocket signaling + WebRTC audio

Sources and verification

Provider specifications describe the upstream model, not a guarantee of every feature through TokenRex. Prices come from the enabled catalog; benchmark results belong to the exact tested model.

Research notes (3)
  • A TokenRex instruction-based mode, not a separately trained OpenAI model or native standalone speech endpoint.
  • Connected time is billed per second with a 15-second minimum; idle time and muted playback remain billable.
  • Exact reading and permanent model silence are not guaranteed. Transcription clients must suppress playback.