Text-to-Speech - Sarj.ai

Sarj TTS converts text into speech using built-in voices or your own reference recording. The speech endpoint supports an OpenAI SDK-compatible request format.

Base URL	Model
`https://sarj-omni-tts.sarj.ai/v1`	`sarj-tts`

Generate speech

Send a JSON request to POST /audio/speech with your TTS API key in the Authorization header. The response is binary audio.

curl --fail-with-body https://sarj-omni-tts.sarj.ai/v1/audio/speech \
  -H "Authorization: Bearer $SARJ_TTS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "sarj-tts",
    "voice": "ars_male",
    "input": "يا مرحبا، كيف نقدر نساعدك اليوم؟",
    "response_format": "wav"
  }' \
  --output speech.wav

Voices and samples

These five Saudi voices use the same sentence, generated with each voice’s production defaults.

يا مرحبا، تقدر تتابع طلبك من التطبيق، وإذا احتجت أي مساعدة حنا موجودين. قل لي وش تحتاج، ونرتبها لك خطوة بخطوة.

Use the ID next to each name as the voice value. For the current voice catalog, call GET /voices relative to the base URL above, with the same API key.

Request parameters

Parameter	Description
`input`	Required. Text to speak, up to 50,000 characters.
`model`	`sarj-tts`, the only supported model and the default.
`voice`	A voice ID from the catalog. Defaults to `auto`; select an ID for a specific voice.
`response_format`	`mp3` (default), `wav`, `opus`, `aac`, `flac`, or `pcm`.
`speed`	Speaking rate, from `0.25` to `4.0`. Omit to use the voice’s configured rate. An explicit value overrides that rate.

For example, add "speed": 1.1 to request a slightly faster delivery than 1.0. Other omitted generation controls use the deployment’s configured defaults.

Advanced controls

Start with the defaults and change one control at a time. t_shift and the two temperatures can affect delivery, but they are not interchangeable expressiveness controls.

Parameter	Effect
`t_shift`	Changes the token-refinement schedule and can affect rhythm, prosody, and stability. Higher values do not necessarily mean more expression. Keep it above `0`; the maximum is `2`.
`class_temperature`	Controls randomness in audio-token selection (`0` to `2`). Higher values add variation but can introduce artifacts; `0` uses greedy selection.
`position_temperature`	Controls randomness in which token positions are refined next (`0` to `10`). It is separate from audio-token selection.
`guidance_scale`	Conditioning strength (`0` to `10`). Increasing it is not a guarantee of better quality.
`num_step`	Number of generation steps (`1` to `64`). Fewer steps can reduce quality; omit to retain the deployment default.
`instructions`	Optional delivery instructions, up to 4,096 characters. Their effect depends on the voice and text.
`language`	Optional language code, such as `ar` or `en`. This does not select a dialect-specific voice.

See the TTS OpenAPI schema for the full request schema, including streaming and additional controls.

Voice cloning

For one-shot cloning, send multipart/form-data to POST /audio/speech/clone. Use a clean reference recording and its exact transcript. A 6-10 second reference is a useful starting point.

curl --fail-with-body https://sarj-omni-tts.sarj.ai/v1/audio/speech/clone \
  -H "Authorization: Bearer $SARJ_TTS_API_KEY" \
  -F "ref_audio=@reference.wav" \
  --form-string "ref_text=The exact words spoken in reference.wav." \
  --form-string "text=يا مرحبا، كيف نقدر نساعدك اليوم؟" \
  -F "response_format=wav" \
  --output cloned-speech.wav

Replace ref_text with the recording’s transcript in its original language. This endpoint uses text, not input, and returns audio without registering a permanent voice.

​Generate speech

​Voices and samples

​Request parameters

​Voice cloning

Generate speech

Voices and samples

Request parameters

Voice cloning