Audio · Synthesis

Text-to-Speech (OpenAI)

A drop-in replacement for OpenAI's /v1/audio/speech: the same request body, the same streamed audio response, spoken in natural Cantonese by our streaming text-to-speech engine. Point an OpenAI SDK at https://cantonese.ai/api/v1 with your cantonese.ai API key and nothing else in your code has to change.

POSThttps://cantonese.ai/api/v1/audio/speech

Field	Behaviour
input	The text to speak (up to 500 characters). Cantonese by default; add `"language": "English"` for English text.
voice	A cantonese.ai `voice_id`, or the name of a public voice. The OpenAI names `alloy`, `echo`, `fable`, `onyx`, `nova` and `shimmer` all exist, so unchanged code keeps working.
response_format	`mp3` (default), `opus`, `aac`, `flac`, `wav` or `pcm` (raw 16-bit little-endian, 24 kHz, mono, the lowest-latency option).
stream_format	`audio` (default, chunked bytes) or `sse` (server-sent `speech.audio.delta` events carrying base64 audio, then `speech.audio.done`).
speed	0.25 to 4.0, default 1.0.
model, instructions	Accepted and ignored: the streaming service serves a single model, V8.2. Ignored fields are listed in the `X-Ignored-Fields` response header.
billing	Characters count 4x against your monthly text-to-speech quota, exactly like the streaming endpoint. An exhausted quota answers `429` with code `insufficient_quota`; the `X-Characters-Charged` header reports what each request deducted.

Errors use OpenAI's envelope ({ "error": { "message", "type", "param", "code" } }): 401 authentication_error for a bad key, 429 rate_limit_error when your plan's per-minute budget is spent, 404 voice_not_found for an unknown voice name.

to auto-fill your API key in the code examples below.

curl -N https://cantonese.ai/api/v1/audio/speech \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"tts-1","voice":"alloy","input":"你好，今日想食啲咩呀？","response_format":"mp3"}' \
  --output speech.mp3