Skip to content

Speech and realtime audio

ASR converts audio to text; TTS converts text to audio. Synchronous files, asynchronous jobs and realtime sessions use different operations and accounting units.

Use caseInterface
OpenAI-compatible file transcription/v1/audio/transcriptions with a supported passthrough or conversion adapter
OpenAI-compatible speech synthesis/v1/audio/speech, with upstream-compatible voice and output format
Native DashScope recognition, synthesis or asynchronous tasksMatching DashScope surfaces; pair submissions and queries using the supported task protocol
Native MiniMax or ElevenLabsImplemented native interfaces and provider request formats
Realtime recognition or synthesisSupported native DashScope / ElevenLabs WebSocket paths

Realtime support does not cover every provider’s speech-to-speech model and does not provide universal OpenAI Realtime conversion.

  1. Enable provider audio endpoints and import an ASR or TTS model.
  2. Choose the adapter for the model family; check file uploads, public audio URL requirements and streaming limits.
  3. Check accounting units. Recognition can use seconds or tokens; synthesis can use characters or tokens, depending on model configuration.
  4. Probe the response in Playground. Supported realtime ASR can use the browser microphone.
  5. Use Simulator or a real client to verify authentication, usage, quotas and logs.

Realtime audio in Playground

Terminal window
curl -sS http://localhost:8787/v1/audio/speech \
-H 'Authorization: Bearer sk-your-api-key' \
-H 'Content-Type: application/json' \
-d '{"model":"your-tts-model","input":"Hello, OctaFuse.","voice":"your-voice"}' \
--output speech.mp3

Use kind=asr or kind=tts for model listing; kind=audio includes both. A provider offering an operation does not by itself mean the gateway implements it.

Read the coverage matrix and user API for exact operations and fields.