POST /v1/speech/synthesize.
Request
Supported languages
Panini supports 15 languages:Voices
The five built-in voices use the same public names in the playground and API:
Omit
voice to use Arjun by default. The backend resolves the selected public
voice to its internal model adapter before contacting the inference server.
When language is provided, the backend forwards it to the model as lang.
When it is omitted, the model detects the spoken language from text.
Pass a built-in slug (e.g.
arjun) as voice, or pass a voice_id from
GET /v1/voices for a saved voice. panini is the model name, not a voice.
To get built-in voice IDs, call GET /v1/voices?type=system; the default
type=all also includes your uploaded voices. Use the slugs in the table
above when sending voice.{"text":"Namaste!","voice":"arjun"} selects Arjun. The playground’s
modelId and outputFormat fields are not part of this API request.
Streaming response (NDJSON)
The endpoint responds withapplication/x-ndjson — one JSON object per line:
Two ways to consume it:
- Stream playback — decode each
chunk_b64as it arrives and play the chunks back-to-back to start audio almost immediately. - Whole file — ignore partials and keep the
is_final: truechunk; it’s the complete WAV, ideal for saving or replay.
Python — save the final WAV
Reference audio
ref_audio is a base64-encoded audio sample for one-off voice cloning. For a
reusable custom voice, upload a WAV or WebM file with POST /v1/voices and
then pass the returned data.voice_id as voice_id when synthesizing.
Built-in voices are sent to the model server as
speaker adapter names.
Uploaded voices and ref_audio use reference audio instead. Choose one
voice selection method per request.