Speech recognition (speech to text, STT) turns the caller's audio into the transcript the model reads. Its settings are under Realtime Transcription Settings. Mistakes here become mistakes in the conversation: a misheard name or number leads to a wrong answer, so tune this section for your callers.
Transcription Mode#
| Option | stt_mode |
What it does | Estimated latency |
|---|---|---|---|
| Optimize for speed | fast (default) |
Deepgram nova-3, transcript finalised 25 ms after speech stops, no smart formatting. |
130-300 ms |
| Optimize for accuracy | accurate |
Deepgram nova-3, finalised after 100 ms, with smart formatting and numerals. Better for names, numbers, emails and addresses. |
300-500 ms |
| Custom | custom |
You pick the provider, the model and the endpointing delay. | 200-450 ms |
Both Deepgram modes stream interim results, keep punctuation and keep filler words.
Custom#
With Custom, a panel appears:
- Provider: Deepgram, AssemblyAI, Soniox, Azure or OpenAI.
- Model: leave empty for the provider default, or type a model name.
- Endpointing: the silence after speech before the transcript is final, 10 to 2,000 ms in the dashboard (the API accepts up to 5,000), default 300 ms. Lower is faster but can cut callers off mid-sentence.
| Provider | Default model | Extra $/min |
|---|---|---|
| Deepgram | nova-3 |
included |
| AssemblyAI | universal streaming (multilingual model for non-English) | 0.004 |
| Soniox | stt-rt-v5 |
0.003 |
| Azure | Azure Speech | 0.025 |
| OpenAI | gpt-4o-mini-transcribe |
0.009 |
Needs keys Each provider needs its key on the platform. When the chosen provider has no key, the call uses Deepgram instead.
Vocabulary#
Vocabulary (vocab_specialization) is General (default) or Medical. Medical uses Deepgram nova-3-medical, a model trained on clinical terms, for English. In multi-language mode it falls back to nova-3.
Languages callers may speak#
Select more than one language to switch transcription to multi-language mode. See Languages.
Denoising Mode#
Noise cancellation runs on the caller's audio before transcription.
| Option | denoising_mode |
Effect |
|---|---|---|
| No denoise | no-denoise |
Audio passes through untouched. |
| Remove noise | noise-cancellation (default) |
Filters background noise. |
| Remove noise + background speech | noise-and-background-speech-cancellation |
Also suppresses other voices, such as a TV or people nearby. |
Noise suppression strength (denoising_enhancement_level) adjusts it:
| Option | Effect |
|---|---|
| Default | The mode above decides. |
| Low | Noise filtering only, even with Remove noise + background speech. |
| Medium | Same as Default. |
| High | Noise and background speech filtering, even with Remove noise. |
With No denoise, the strength is ignored. Phone calls use a background-voice model tuned for telephone audio.
Use Remove noise + background speech for callers in cars, shops or open-plan offices, where nearby voices otherwise get transcribed. Higher strength can clip very quiet speech, so test with real recordings.
Needs keys Denoising uses LiveKit's noise cancellation and needs LiveKit Cloud credentials.
Boosted Keywords#
Boosted Keywords (boosted_keywords) lists words the transcriber should favour: brand names, product names, surnames, suburbs and jargon. Type a word and press Enter.
- Keywords can contain dynamic variables, for example
{{first_name}}, which are filled in per call. - Deepgram receives them as key terms. AssemblyAI receives them as a key term prompt.
- Keep the list focused. Boosting common words can make the transcriber hear them where they were not said.
Fields#
| Field | Type | Default |
|---|---|---|
stt_mode |
fast, accurate, custom |
fast |
custom_stt_config.provider |
deepgram, assemblyai, soniox, azure, openai |
deepgram |
custom_stt_config.model |
string | provider default |
custom_stt_config.endpointing_ms |
integer 10 to 5000 | 300 |
vocab_specialization |
general, medical |
not set (general) |
languages |
string array | not set |
denoising_mode |
enum above | noise-cancellation |
denoising_enhancement_level |
low, medium, high |
not set |
boosted_keywords |
string array | [] |
{
"stt_mode": "custom",
"custom_stt_config": { "provider": "assemblyai", "endpointing_ms": 400 },
"denoising_mode": "noise-and-background-speech-cancellation",
"boosted_keywords": ["Harbourside", "Invisalign", "{{first_name}}"]
}Per-node overrides#
Conversation flow nodes can override stt_mode and boosted_keywords while the node is active, for example accuracy mode only on the node that collects an email address. See Conversation flow.
Workspace fallback#
Under Settings, Reliability, Speech recognition fallback names a second provider used when the agent's transcription provider fails. See Reliability.
Data residency#
The platform operator can point Deepgram at a regional endpoint, for example Deepgram's AU endpoint, so Australian audio is transcribed in Australia.
Check transcription quality#
- Read the transcript in Call History next to the recording.
- Per-word timestamps are stored in
transcript_object[].words. - AI Quality Assurance can measure word error rate (WER) on calls you correct. See Analysis.