Getting LiveKit TTS returned error: Provider deepgram error (status 403): Authentication/authorization error - retryable: false but only with deepgram TTS provider using it as instructed in the docs

tts: new inference.TTS({
  model: "deepgram/aura-2",
  voice: "apollo",
  language: "en",
}),

Thanks for reaching out, Karan. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

Inference is on, it works for some providers like elevenlabs/inworld but not any of the others

Thanks — that helps. I’m responding to your note that Inference works for some providers (ElevenLabs, Inworld) but Deepgram TTS returns a 403. Because other providers work, we can rule out a global LiveKit credential or Inference-disabled project problem. The likely causes to check next are small configuration differences for Deepgram or a provider-specific access/billing restriction. Please try these targeted checks and share the results: 1. Confirm the exact model/voice string you’re sending ◦ Use the documented model/voice format for Deepgram:

model: "deepgram/aura-2"

and a supported voice such as

"athena"

(some docs show

athena

for examples). If you used

apollo

in your original snippet, try

athena

or the exact voice name documented for

aura-2

. 1. Try the minimal reproducible snippet and capture the full error ◦ Run the same code in a minimal environment and paste the full error object/response (not just the message). The full JSON or server log often includes a provider error code or reason field that points to whether it’s a permission, region, or billing denial. 1. Confirm language/voice combinations are supported ◦ Deepgram TTS via Inference supports

en

and

es

. If you requested a voice/language combo not supported by the chosen model, that could trigger an authorization-like denial. 1. Compare request differences between a working provider and Deepgram ◦ Compare headers, model strings, and any additional fields you send. If you’re programmatically switching providers, a small typo or extra field might cause the Deepgram path to fail. 1. Check project billing/quotas with provider-level granularity ◦ Your project-level billing can still block specific downstream providers in some account setups. Confirm there aren’t provider-specific usage restrictions on Deepgram for your LiveKit project. If you can, please provide: • The exact

inference.TTS

call you’re running (copy/paste). • The full error object or server log for the 403. • The environment (local, Docker, serverless) where you’re running this. With those details I can point to the precise next step — for example, whether to change the voice string, adjust language, or inspect a particular error field in the logs. Sources: LiveKit Inference | LiveKit Documentation | Deepgram STT | LiveKit Documentation | Text-to-speech (TTS) models overview | LiveKit Documentation | Deepgram TTS | LiveKit Documentation, with interactive elements