Models & Languages

Oogam exposes every capability behind a stable model id. Pass the id in the model field of a request. These four are the whole catalogue — voice and speech pricing also comes back live from GET /voices.

Model catalogue

Everything is on v2. To upgrade an existing request, change only its model value — your API key and the endpoint URLs are unchanged. Send model on every request: omitting it still returns audio, from whichever model is configured as the default, which is not a guarantee you want to build on.

Voice work uses the naad-* models above (see the Audio & Voice reference); sutra-conversation serves Text & extraction on /chat/completions, for keys with the Sutra product enabled.

naad-tts-v1, naad-stt-v1 and naad-sts-v1 remain accepted for existing integrations and return the same output as their v2 ids. They are not advertised: GET /models lists the four above.

Not in the catalogue. naad-text-v1 is a legacy id for the same model as sutra-conversation — it works, but write sutra-conversation in new code. There is no embeddings model: naad-embed-v1 returns 503.

The conversational model

sutra-conversation is the only LLM we publish, and it is ours: it runs on our own GPUs, so nothing you send is resold to a third-party provider. It covers all 27 Indian languages in-house.

It will not state a bill amount, balance, due date, payment status or account status. This is deliberate, not a defect: the model cannot see the caller's account, and when it used to guess it invented figures in 10 of 81 replies — including an impossible “35th of next month”. It now asks for the account number instead. If you need real figures spoken, supply them through a grounded session (set_context or a knowledge base); grounded sessions are exempt from this rule by design.

Replies come back in the caller's native script and run about 15 words. In the realtime pipeline the model also emits a leading (emotion) tag, which the gateway consumes and strips for you. Calling /chat/completions directly, we did not see the tag on plain chat turns — but it is produced under the realtime prompt, so strip a leading (word) tag defensively rather than assuming it is absent.

List models — GET /models

Returns the catalogue in an OpenAI-compatible shape, so standard SDK bootstrap calls work. The live flag tells you which models are reachable on this deployment. Only usable models are listed — anything whose provider is not configured is left out rather than returned with live: false, so an SDK that iterates the array cannot land on a model that answers 503. Every row you get back has live: true.

Superseded ids are not listed, but they still resolve when sent, so an older integration keeps working without a change.

curl https://platform.oogam.ai/v1/models \
  -H "Authorization: Bearer $OOGAM_API_KEY"

{
  "object": "list",
  "data": [
    { "id": "sutra-conversation", "object": "model", "kind": "chat", "owned_by": "oogam", "live": true },
    { "id": "naad-tts-v2", "object": "model", "kind": "tts", "owned_by": "oogam", "live": true },
    { "id": "naad-stt-v2", "object": "model", "kind": "stt", "owned_by": "oogam", "live": true },
    { "id": "naad-sts-v2", "object": "model", "kind": "sts", "owned_by": "oogam", "live": true }
  ]
}

Language codes

Every language we serve is listed below (29 in all, 27 of them Indian). TTS and STT take a 3-letter code; the realtime WebSocket (Live talk) takes the short code. They match for every language except Indian English. The models never auto-detect — always send the spoken language.

Tier is the quality signal: Lead is our flagship quality, Beta is live at lower accuracy, and Preview is available but not yet at flagship quality. Every tier is callable on the same endpoints and billed the same; the tier tells you what to expect, not what you may use.

Indian English is ENG on /audio/speech and /audio/transcriptions, and en-IN on the realtime socket. Sending en-IN to the batch endpoints also works — it resolves to the same voice and model — but ENG is the code to write down. IEN is a retired spelling: TTS still accepts it, transcription refuses it with 400 missing_language.