Beta

API reference

Three endpoints, OpenAI-compatible where it counts. Read the whole contract before you decide whether to sign up.

Base URL and authentication

Every request carries an API key as a bearer token. Create an account to get one; it comes with ⁦$5⁩ of credit and needs no card. The speech and transcription API endpoints require a valid API key.

https://api.sawtakarabi.ai/v1

Authorization: Bearer sk_ar_live_...

The speech and transcription endpoints are OpenAI-compatible. Point an official OpenAI SDK’s baseURL here and use your Sawtak Arabi key.

Text to Speech

POST /v1/audio/speech

modelstring

Use arabic tts with OpenAI SDKs.

inputstring

The text to speak. Billed per character.

voicestring

A voice id from your Voices page. Omit it for the default zero-shot voice.

response_formatstring

Must be "pcm". Raw 16-bit mono PCM is the only format we emit.

sample_ratenumber

Optional output rate from 8000 to 48000 Hz. Defaults to 24000 Hz.

Field	Type	Notes
model	string	Use arabic tts with OpenAI SDKs.
input	string	The text to speak. Billed per character.
voice	string	A voice id from your Voices page. Omit it for the default zero-shot voice.
response_format	string	Must be "pcm". Raw 16-bit mono PCM is the only format we emit.
sample_rate	number	Optional output rate from 8000 to 48000 Hz. Defaults to 24000 Hz.

curl --fail --show-error https://api.sawtakarabi.ai/v1/audio/speech \
  -H "Authorization: Bearer sk_ar_live_..." \
  -H "Content-Type: application/json" \
  -d '{"model":"arabic tts","input":"مرحبا بك","voice":"9287b630-bc5f-574d-8c38-5f525c2b4ca3","response_format":"pcm"}' \
  --output speech.pcm

The response body is raw little-endian 16-bit mono PCM, with no WAV header. The rate it was actually emitted at comes back on X-Sample-Rate; read that header rather than assuming a rate, or the audio plays at the wrong speed with no error anywhere.

Speech to Text

POST /v1/audio/transcriptions

modelstring

Use arabic asr. Required by OpenAI SDKs.

filemultipart

Audio or video file. The complete multipart upload must fit within 25 MB.

response_formatstring

Optional: json (default, recommended) or text. verbose_json is also accepted but does not add word or segment timestamps.

languagestring

Optional BCP-47 language code. Defaults to ar.

temperaturenumber

Optional sampling temperature. Defaults to 0.

Field	Type	Notes
model	string	Use arabic asr. Required by OpenAI SDKs.
file	multipart	Audio or video file. The complete multipart upload must fit within 25 MB.
response_format	string	Optional: json (default, recommended) or text. verbose_json is also accepted but does not add word or segment timestamps.
language	string	Optional BCP-47 language code. Defaults to ar.
temperature	number	Optional sampling temperature. Defaults to 0.

Prepare a recording with clear Arabic speech. Replace meeting.wav in the example with your local file path. Upload the file as multipart form data; do not send a JSON body or manually set the multipart boundary.

curl --fail --show-error https://api.sawtakarabi.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer sk_ar_live_..." \
  -F model="arabic asr" \
  -F [email protected] \
  -F response_format=json

Example JSON response

{
  "text": "النص المكتوب",
  "usage": { "type": "duration", "seconds": 42 }
}

Transcription is charged after the audio is processed, on usage.seconds, because the duration is not knowable until the file is decoded. Your balance has to cover the longest the upload could be before the request starts.

The json response includes the transcript in text and usage when supplied by the transcription service. With response_format=text, the body is plain text and has no usage object. Neither format provides speaker labels or word timestamps.

Review names, numbers and specialist terms before publishing or using a transcript in an important workflow. Audio quality, overlapping speakers and mixed languages can affect the result.

Voices

Browse the Voices page, open a voice and copy its Voice ID. Pass that exact value as voice; the display name and dialect label are not identifiers.

Your API key can use voices owned by its account and voices published for public use. Knowing a private voice ID does not grant access. A deleted, unavailable or unauthorized voice returns an error; the API does not silently substitute another voice.

To create a voice from your recording, follow the voice cloning guide. Once encoding finishes, use the new voice ID in the same speech request.

Limits

requests600 / min

Default per account, shared across its API keys.

concurrency5

Default simultaneous in-flight requests per account.

upload25 MB

Maximum transcription request body, including multipart overhead.

Field	Type	Notes
requests	600 / min	Default per account, shared across its API keys.
concurrency	5	Default simultaneous in-flight requests per account.
upload	25 MB	Maximum transcription request body, including multipart overhead.

Errors

Read the HTTP status first. Gateway errors normally include a JSON detail message, X-Error-Code for the machine-readable reason, and X-Request-Id to identify the request. Errors from a proxy or interrupted connection may not include these fields.

Status	Meaning
400	Malformed body, unsupported response_format, or an unusable voice value.
401	Missing, revoked or unrecognised API key.
402	Not enough prepaid balance for the request.
404	The voice id does not exist or is not available to your account.
413	The request body exceeds the endpoint limit. Reduce the upload or input size before retrying.
429	Rate limit or concurrency limit. Retry with backoff.
502	The gateway could not complete the request with the speech service. Wait before retrying.
503	No capacity right now. Retry shortly.

We refund requests that fail on our side before producing audio. Invalid options and unavailable voices are rejected before charging. A backend rejection of your request after charging, a client disconnect, or a partially delivered stream is not refunded.

Fix input, credentials or balance before retrying 400, 401, 402, 404 or 413 responses. For 429, 502 or 503, retry a limited number of times with increasing delays. Avoid parallel retry loops. If a connection breaks after submission, the outcome can be uncertain; another request can generate and charge again.

When contacting support, include the request ID, endpoint, HTTP status and approximate time. Do not include your API key or the recording itself.

What it costs

Text to Speech: ⁦$50⁩ / 1M characters
Speech to Text: ⁦$0.50⁩ / hour of audio

Prepaid. At a zero balance, paid API requests and calls stop rather than running up an invoice. start with free credit to resume.