Speech-to-text (STT) turns the caller's audio into text for the LLM. The default Edesy runtime supports Deepgram (Nova and Flux), Google Chirp, Azure, AssemblyAI, ElevenLabs Scribe, Sarvam and Soniox. Deepgram is the default and is included in the per-minute price. For Indian languages and mixed Hindi-English speech, test Sarvam, Soniox or Google.
STT only applies to pipeline agents (STT + LLM + TTS). Native audio agents such as Gemini Live listen directly and do not need a separate STT.
Supported providers#
| Provider | Provider ID | Status | Best for |
|---|---|---|---|
| Deepgram Nova | deepgram |
Default runtime, default provider | English and general production agents |
| Deepgram Flux | deepgram_flux |
Default runtime | Model-decided end of turn, fewer cut-offs |
| Google Chirp | google |
Default runtime | Broad language coverage, Indian languages |
| Azure Speech | azure |
Default runtime | Enterprise Azure setups, Indian locales |
| AssemblyAI | assemblyai |
Default runtime | English accuracy |
| ElevenLabs Scribe | elevenlabs |
Default runtime | Regional languages such as Assamese |
| Sarvam | sarvam |
Default runtime | Indian languages and code-mixed speech |
| Soniox | soniox |
Default runtime | Malayalam, Tamil, Bengali and multilingual callers |
| Reverie | reverie |
Available on request | Indian languages, set up with our team |
| OpenAI Whisper | openai_whisper |
Available on request | Batch-style multilingual transcription |
If an agent names a provider the runtime does not support, the call falls back to Deepgram and the substitution is logged. Pick a provider from the table above to avoid that.
Choosing a provider#
| Your callers speak | Start with | Also test |
|---|---|---|
| English | Deepgram Nova | Deepgram Flux, AssemblyAI |
| Hindi or Hinglish | Sarvam (codemix mode) |
Deepgram, Google |
| Tamil, Telugu, Malayalam, Kannada, Bengali | Soniox | Sarvam, Google |
| Gujarati, Marathi, Punjabi, Odia | Sarvam | Google, Soniox |
| Assamese | ElevenLabs Scribe | Soniox |
| Many languages on one line | Soniox (language identification) |
Edesy supports 75+ languages overall, including 14 Indian languages and Hinglish and bilingual modes. See language support.
How to switch provider#
- Open the agent and go to Voice → Speech-to-text.
- Choose the provider and model.
- Save, then place a test call from the Test Center.
- Open the call in call history and read the transcript. Compare two or three providers on the same script before you decide.
Through the API, set sttProvider (and optional sttConfig) on the agent:
{
"sttProvider": "sarvam",
"sttConfig": { "model": "saaras:v3", "mode": "codemix" }
}
Pricing#
Deepgram STT is included in the per-minute price. Some providers add a per-minute charge (for example Google ₹1.00/min and Azure ₹0.80/min). See provider costs for the current list. With your own keys, see BYOK.
FAQs#
Which STT provider is the default? Deepgram Nova.
Does the STT choice matter for native audio agents? No. Gemini Live and OpenAI Realtime handle speech themselves.
Can I bring my own STT key? Yes, for Deepgram, Sarvam, Soniox, AssemblyAI, ElevenLabs, Azure and others. See BYOK.
Are Speechmatics or other providers supported? Only the providers in the table above run on the default runtime.