Speech-to-Text

Speech-to-text (STT) transcribes audio into text. You send the audio inline as base64 inside the JSON body — there is no multipart upload — and the model returns a transcript.

Endpoint#

POST https://opendunes.com/api/v1/audio/transcriptions

Request body#

Audio goes under input_audio, either inline as base64 plus its container format, or as a reference to a file you already uploaded to the Files API. Send one or the other, not both.

curl https://opendunes.com/api/v1/audio/transcriptions \
  -H "Authorization: Bearer $OPENDUNES_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "example-provider/stt-model",
    "input_audio": { "data": "'"$(base64 -i clip.mp3)"'", "format": "mp3" },
    "language": "ar"
  }'

Transcribing the same clip more than once? Upload it and reference it by id — the bytes travel over the wire once instead of on every call:

FILE_ID=$(curl -s https://opendunes.com/api/v1/files \
  -H "Authorization: Bearer $OPENDUNES_API_KEY" \
  -F file=@clip.mp3 | jq -r .id)

curl https://opendunes.com/api/v1/audio/transcriptions \
  -H "Authorization: Bearer $OPENDUNES_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{
    \"model\": \"example-provider/stt-model\",
    \"input_audio\": { \"file_id\": \"$FILE_ID\" }
  }"

Response#

{
  "text": "أهلاً وسهلاً! هذا اختبار لنظام التعرف على الكلام."
}

Language support#

Arabic (Modern Standard Arabic and Algerian Darija), French, and English are priority targets. Darija accuracy depends on the underlying model — check the model detail page for supported languages.

Pricing#

STT is billed in DA per second of audio processed (audio seconds are rounded up). Rates vary by model — see the pricing reference and the model catalog.

Finding STT models#

GET https://opendunes.com/api/models?modality=transcription

Related#

• Text-to-Speech — generate spoken audio from text
• Audio Inputs — send audio to multimodal language models