Transcriber (STT)
Deepgram
Deepgram Nova-3 is the recommended STT provider for production deployments — best-in-class accuracy with an average latency of 83ms and 37 language codes including 15+ Arabic dialect variants.
Requires: Deepgram API key in Settings → Credentials
Models
| Model | Use Case |
|---|---|
nova-3 | Recommended — highest accuracy |
nova-3-medical | Healthcare terminology |
Configuration
Code
Fields
| Field | Default | Description |
|---|---|---|
model | nova-3 | STT model |
language | en-US | Language code (see list below) |
detect_language | false | Auto-detect language (disables language) |
smart_format | true | Format numbers, dates, addresses |
punctuate | true | Add punctuation |
numerals | false | Convert spoken numbers to digits |
enable_diarization | false | Identify different speakers |
keyterms | null | Terms to boost recognition (Nova-3 only) |
interim_results | true | Return partial transcripts during streaming |
endpointing_ms | 25 | Silence before utterance end detection (10–5000ms) |
filler_words | true | Include "um", "uh" in transcripts |
profanity_filter | false | Filter profanity from output |
mip_opt_out | false | Opt out of Deepgram's model improvement program |
Language Codes
37 supported codes:
English: en, en-US, en-GB, en-AU
Arabic (15+ dialects): ar, ar-AE, ar-SA, ar-EG, ar-QA, ar-KW, ar-JO, ar-LB, ar-SY, ar-IQ, ar-MA, ar-DZ, ar-TN, ar-PS, ar-SD, ar-TD, ar-IR
Other: es, es-419, fr, de, it, pt, pt-BR, nl, hi, ru, ja, zh, ko, tr, uk, pl, vi
Multilingual (any language): multi
Keyterms (Nova-3)
Boost recognition of domain-specific terms:
Code
Terms are case-insensitive. List up to 100 keyterms per agent.
Related
- OpenAI STT — Whisper-based alternative
- Phantom STT — coming soon
Last modified on
