Models
Models
We provide API endpoints for our open models: oolel-speech-v1 and oolel-voices.
Oolel Speech V1
Recommended prompts and hyperparameters
Use these settings as your starting point. For speech recognition,
temperatureshould be0.0; this is the most important setting on this page because recognition should be deterministic.
Speech input length: For optimal Oolel Speech results, keep speech inputs shorter than 30 seconds. Split longer recordings before sending them to the model.
Recommended Speech Settings
| Task | Recommended prompt | Recommended temperature |
|---|---|---|
| Speech recognition | | 0.0 |
| Speech translation | |
|
| Text translation | |
|
For the full accepted range of each parameter, visit the API Reference.
The first version of Oolel Speech (oolel-speech-v1) is a Speech LLM optimized for text translation, speech translation, and speech recognition.
For text and speech translation, the model supports bidirectional Wolof (wo) to/from English (en) and Wolof (wo) to/from French (fr). The model supports speech recognition for Wolof, English, and French.
The dedicated text translation APIs also accept all three target codes, including
target_language: "wo" for English or French input.
Important: The Speech LLM is optimized mainly for speech and text translation, and speech recognition. The model is not built for chat tasks.
These examples use the OpenAI-compatible Chat Completions route:
Code
Set your API key first:
Code
The examples below use these recommended prompts and starting hyperparameters. Keep the instruction direct and ask for the target output only.
Text Translation
Translate Wolof text to French.
Speech Translation
Translate speech audio to French text. The curl example writes the request body to a file so large base64 audio does not exceed your shell's command length limit.
Speech Recognition
Transcribe speech audio to text.
Required recommendation: Use
temperature: 0.0for speech recognition. This keeps transcription deterministic and reduces avoidable variation in the recognized text. The examples below intentionally settemperatureto0.0.
Oolel Voices
Recommended voice parameters
Use these ranges as a starting point, then adjust them by listening to generated samples for your own voice, text length, and use case.
Recommended Voice Settings
| Parameter | Recommended value | Role |
|---|---|---|
exaggeration | 0.1 to 0.5 | Controls expressiveness and emphasis. Start near 0.1 for a more neutral delivery and increase toward 0.5 for more expressive speech. Higher values may work particularly well for some English-language content. |
temperature | 0.1 to 0.5 | Controls generation variability. Lower values produce more consistent output, while higher values introduce more variation. Test within this range for your voice, language, and use case. |
cfg_weight | 0.5 | Controls how strongly the generation follows the text and conditioning. Higher values can enforce the prompt more strongly; lower values can sound more relaxed. |
seed | 0 | Makes repeated tests easier to compare. Change it when you want a different generation with the same other settings. |
reference_audio_url | Optional | Uses hosted reference audio to guide voice style. Include the voice-rights fields when you use reference audio. |
voice_rights_attested | true for reference audio | Confirms you have the rights to use the reference voice sample. |
voice_rights_basis | self or authorized | Describes why you are allowed to use the reference voice sample. |
For a more expressive preset, try exaggeration: 0.5, temperature: 0.5, and
cfg_weight: 0.5. This can work particularly well for some content in English
or French, but compare results for your target voice and text.
For the full accepted range of each parameter, visit the API Reference.
Oolel Voices generates speech from text. Omit reference audio to use Soynade's default Oolel Voices speaker. Add reference_audio, reference_audio_url, or reference_audio_base64 only when you want to guide the voice style from a sample you have permission to use.
Every reference-audio request must include voice_rights_attested=true and a voice_rights_basis of self or authorized. See Voice Rights and Safety.
Direct Speech Generation
Create a WAV file directly.
Reference Audio URL
Use hosted reference audio for voice style.
Local Reference Audio
Use a local reference audio file for voice style.
Use reference_audio_base64 only when file upload and hosted URLs are not practical. New clients should prefer reference_audio_url or multipart reference_audio for larger files. The same attestation fields are required for base64 reference audio.
Async Speech Jobs
For longer text or production workflows, create an async job, check its status, then download the generated audio when the job is complete.
Submit A Job
The response includes a job_id. Use that value in the status and download requests.
Check Job Status
Wait until the job status is completed before downloading the audio.