Provider reference

Choose an adapter by capability, dependency, and credential. A configured route tries providers in order. Each provider still has its own language and model limits, so consult its guide before depending on a fallback.

Provider ID	Translation	Detection	Transliteration	STT	TTS	Setup
`bhashini`	Yes	Yes	Yes	Yes	Yes	API key, endpoint, capability service IDs
`sarvam`	Yes	Yes	No	Yes	Yes	`SARVAM_API_KEY`; STT/TTS model IDs
`navana`	No	No	No	No	Yes	`NAVANA_API_KEY`; HTTP or WebSocket TTS
`gnani`	No	No	No	Yes	Yes	`GNANI_API_KEY`; streaming and non-streaming STT/TTS
`googletrans`	Yes	No	No	No	No	`[googletrans]` extra; unofficial online adapter
`fasttext`	No	Yes	No	No	No	`[local-tld]` extra; local model
`aksharamukha`	No	No	Yes	No	No	`[local-transliteration]` extra
`indicxlit`	No	No	Yes	No	No	Adapter retained; install extra omitted from beta
`google_free`	No	No	No	Yes	No	`[stt-google-free]` extra; unofficial online adapter
`faster_whisper`	No	No	No	Yes	No	`[stt-whisper]` extra; local model
`edge_tts`	No	No	No	No	Yes	`[tts-edge]` extra; unofficial online adapter

Routing and errors

Set route order under [routes] in .indic-language-utils.toml. For example:

[routes]
text_to_speech = ["sarvam", "bhashini", "edge_tts"]

The client selects providers whose declarations support the requested language. It retries or falls back on rate limits, timeouts, transient provider errors, malformed responses, and output validation failures. Authentication, invalid input, and unsupported language errors stop the request. For TTS with no language, Sarvam is skipped because it requires one. Bhashini can handle an unspecified language only when it has a default TTS model. Navana uses Hindi when language is omitted.

TTS options are provider-specific. TTSOptions.parameters applies to the first compatible provider only. Use TTSOptions(provider_parameters={...}) when fallback providers need their own voice settings. See the TTS guide for an example.

Before integrating

Install only the extras your route needs. Local model adapters may download or load model files on first use.
Verify a provider's supported languages and its model version. See the Models and voices reference for a catalog of model IDs, service IDs, neural voices, and parameters.
Use async with get_translation_client(), get_stt_client(), or the matching client factory when making repeated calls. This closes network resources after use.
Inspect result.provider, model_id, and fallback_count where available to see which engine produced the result.
The demo workbench's web assets are built from a source checkout. They are not part of the Python wheel.