Text to speech

POST /v1/audio/speech turns text into audio for narration, voice replies, or notifications. The response body is the raw audio, with no JSON wrapper; Content-Type names its format.

curl https://api-gateway.merge.dev/v1/audio/speech \
  -H "Authorization: Bearer $MERGE_GATEWAY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "openai/tts-1", "input": "Your order has shipped and arrives Thursday.", "voice": "alloy"}' \
  -o speech.mp3

Voices

Model	Voices	Output
`openai/tts-1`, `openai/tts-1-hd`	OpenAI voices such as `alloy` and `nova`	`response_format`, MP3 by default
`google/gemini-3.8-flash-tts`, `google/gemini-3.8-flash-lite-tts`	Gemini voices such as `Kore` or `Puck`, or omit `voice`. OpenAI names fail with `No matching speaker voice found`.	WAV, whatever `response_format` says
`bland/bland-speech-v3`	OpenAI voice names, mapped to Bland voices, or a Bland voice UUID	WAV
Other vendors, such as `minimax/speech-2.6-turbo`	The vendor’s own voice IDs	Vendor-specific

The Gemini TTS models run on the google vendor only and take up to 8,192 input tokens.

Reference

Field	Notes
`model`	Required. A canonical model ID, since `@alias/...` returns `400 alias_not_supported`.
`input`	Required. The text to speak.
`voice`	Model-specific, see above
`response_format`	`mp3`, `opus`, `aac`, `flac`, `wav`, or `pcm`
`speed`	0.25 to 4.0
`instructions`	Voice direction, on models that support it
`vendor`	Pin the execution vendor
`customer`	Customer UUID to scope the key, budget, and usage to

Other fields are forwarded to the vendor. Speech bills per input character (unit: per_character on GET /v1/models). X-Merge-Billed-Characters gives the count billed after any DLP redaction, and X-Merge-Vendor names the vendor.

Next steps

Speech to text

Transcribe an audio file

Live voice

Full-duplex voice sessions over a WebSocket