Skip to main content

Models

Djelia model names clearly identify each public capability family. The number following the dash indicates the generation of that family.

ModelCapabilityEndpointStatus
sunjata-1Bambara speech-to-text/audio/transcriptions, /audio/translationsCurrent
jifili-1Bambara text-to-speech/audio/speechCurrent
banjugu-1Bambara, French, and English text translation/chat/completionsCurrent
djelia-asr-v1Bambara speech recognition, first generation/audio/transcriptions, /audio/translationsAvailable until September 19, 2026
djelia-asr-v2Bambara speech recognition, optimized for short audio with French content/audio/transcriptions, /audio/translationsAvailable until September 19, 2026
djelia-tts-v1Bambara speech synthesis with fixed speaker IDs/audio/speechAvailable until September 19, 2026
djelia-tts-v2Bambara speech synthesis with prompt-steerable voices/audio/speechAvailable until September 19, 2026

The aliases djelia-asr-v3, djelia-tts-v3, and djelia-translate-v1 resolve to sunjata-1, jifili-1, and banjugu-1 respectively.

To view the models accessible with your key, use GET /openai/v1/models or client.models.list().

OpenAI Model Aliases

Djelia accepts OpenAI model names as aliases, enabling most integrations to work seamlessly—just swap out the base URL.

AliasResolves to
whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribesunjata-1
tts-1, tts-1-hd, gpt-4o-mini-ttsjifili-1

All responses return the canonical Djelia model ID for complete transparency—there’s no risk of misidentifying a model as OpenAI’s.

sunjata-1

sunjata-1 excels at transcribing Bambara audio. Calling /audio/transcriptions provides results in Bambara, while /audio/translations returns English by default or French if you set language=fra_Latn.

Supported transcription response formats include: json, text, verbose_json, srt, and vtt. You can stream with json and text. The temperature field is supported for sampling.

jifili-1

Use jifili-1 to generate Bambara speech. The current default voice is moussa.

Output formats supported: mp3, wav, pcm, opus, ulaw, alaw, l16_8000, l16_16000, and fmp4. For streamed audio, set stream_format="sse" to receive speech.audio.delta events; otherwise, you receive audio bytes.

Sampling parameters can be tuned via djelia.temperature, djelia.top_p, and djelia.repetition_penalty. More details are available under Djelia extensions.

Legacy Voices

Until September 19, 2026, djelia-tts-v2 includes three voices:

  • moussa
  • sekou (calm and measured)
  • seydou (bright and energetic)

OpenAI voice names map to Djelia voices as follows:

OpenAI voiceDjelia voice
alloy, nova, coralmoussa
echo, onyx, ash, sagesekou
fable, shimmer, ballad, verseseydou

jifili-1 provides the moussa voice.

banjugu-1

banjugu-1 enables seamless translation via /chat/completions. Provide the text to translate in the last user message and specify the language pair using djelia.source_language and djelia.target_language.

Languages

Djelia adheres to the FLORES-200 language codes.

CodeLanguage
bam_LatnBambara (Bamanankan)
fra_LatnFrench
eng_LatnEnglish

Translation usage is counted in characters. Streamed translations are delivered as a single content delta, since the model generates the entire translation before responding.

Previous Model Names

The legacy models djelia-asr-v1, djelia-asr-v2, djelia-tts-v1, and djelia-tts-v2 are being phased out. They will be supported until September 19, 2026, after which they will return 404 errors.

Aliases like djelia-asr-v3, djelia-tts-v3, djelia-translate-v1, and the OpenAI model names now point to sunjata-1, jifili-1, and banjugu-1, inheriting the current behavior. For complete details on model mapping and changes, refer to the model migration guide.