STT Settings
Use the stt object in an Advanced Settings resource to bias recognition toward specific words and tune transcript finalisation after silence.
Overview
With supported providers, keyterms helps recognise names and domain-specific vocabulary. endpointing controls how long the transcription engine waits after silence before finalising a transcript.
STT settings work alongside VAD Settings and the turn detector. Transcript finalisation and end-of-turn detection run together, and the platform can start generating a response before the turn is committed. Provider buffering, turn detection and response generation all affect when the caller hears a reply; the two configured delays are not a complete latency formula.
Note
STS (Speech-to-Speech) models consume audio directly and have no separate transcription step, so endpointing does not apply to them. keyterms still applies, through a different mechanism. See STT settings with STS models below.
Data model
Both fields are optional. Omit keyterms to use an empty list. Omit endpointing or set it to null to use the integration default.
| Field | Type | Default | Description |
|---|---|---|---|
keyterms | string[] | [] | Individual words the STT engine should prioritise |
endpointing | number | null | null (integration default) | Seconds of silence before the transcript is finalised |
Example:
Code
Provider support
Not every STT provider supports every setting. For endpointing: null, the platform uses the integration defaults below. These values are sent by the integration and may differ from the provider's native defaults.
| Model | keyterms | endpointing | ID |
|---|---|---|---|
| Deepgram Nova-3 | Yes | Yes (default: 0.025s) | 0dda8a0b-5a55-42dc-9071-88b75bff5d78 |
| Deepgram Nova-2 | Yes | Yes (default: 0.025s) | 7a67c6a2-c2a1-40ae-ad78-720552cf2482 |
| AssemblyAI Universal-3.5 Pro Realtime | Yes | Yes (default: 0.1s) | a7c3e9f1-4b2d-4e8c-9a5f-1d7b3e6c2a01 |
| Gladia Solaria-1 | Yes | Yes (default: 0.05s) | 56510288-e114-4bed-9357-3fb57188a54c |
| Speechmatics Standard | Yes | Yes (default: 0.5s) | f5b82a3d-9c4e-4d0f-b2a6-7e3f9c5d8b01 |
| Speechmatics Enhanced | Yes | Yes (default: 0.5s) | f5b82a3d-9c4e-4d0f-b2a6-7e3f9c5d8b02 |
| Voxtral Mini Transcribe Realtime | Yes | No | 600592b5-54d1-4bc4-a8ab-770c376b235f |
| Voxtral Mini Transcribe 2 | Yes | No | 3da23fc4-ffd8-4e19-8145-e788317ca27c |
| GPT Realtime 2.1 | Yes | No | e5a9c8d6-0f7b-4e2c-9d0a-6b2f4c0e7d08 |
| GPT Realtime 2.1 Mini | Yes | No | f6b0d9e7-1a8c-4f3d-8e1b-7c3a5d1f8e09 |
| GPT Realtime 2 | Yes | No | b3d34d2a-8254-405a-8704-4108b2f097d7 |
| GPT Realtime 1.5 | Yes | No | b1d67e16-b522-49f4-88e9-6f0acac76dbb |
For Deepgram Flux, endpointing is mapped to eot_timeout_ms and clamped to 0.5 to 60 seconds. The Cartesia, Gradium and Voxist integrations do not apply these STT tuning fields. LiveKit inference does not receive the configured endpointing; its support for session keyterms depends on the underlying model.
The GPT Realtime entries are STS models rather than transcription engines. They accept keyterms through the mechanism described below, and never endpointing.
STT Settings with STS models
An STS model receives audio and produces audio without a separate transcription step, which changes how the two stt fields behave.
endpointing has no effect on STS models. Use the supported vad fields to configure turn-taking. See VAD with STS models for provider support.
keyterms is added to an STS Assistant's instructions as a dedicated section. It tells the model that these terms may appear and to spell them exactly as written. A separate transcription engine instead receives keyterms as recognition hints.
The effect of instruction-based keyterms depends on the STS model. Use the spellings you want the Assistant to produce. Recognition is not guaranteed.
Note
Keyterms use prompt space on an STS model. Include terms that are often misheard instead of the full product catalogue.
Configuring STT properties
keyterms
List words the STT engine should prioritise during recognition, such as proper nouns, brand names, product names, or domain-specific vocabulary. Multi-word phrases are accepted by the API; support depends on the provider. The Voxtral integration splits entries on whitespace and commas into individual tokens.
Code
Provide the spelling you want the recogniser to favour. For example, add "Reecall" if it is often transcribed as "recall". Keyterms are hints and do not guarantee a particular transcription.
Keyterms are capped at 100 entries and 100 characters per entry. Empty or whitespace-only strings are rejected.
endpointing
Controls how long (in seconds) the STT engine waits after detecting silence before finalising the transcript. This is a separate silence timer from the VAD's endpointingMinDelay: VAD decides when the user's turn ends at the audio level, while endpointing determines when the STT engine commits its transcript.
Code
The STT engine and turn detector both affect response timing, but buffering, adaptive endpointing and preemptive generation mean the audible delay cannot be calculated as the maximum of these two settings. Test both controls with representative calls.
- Lower values (
0.2to0.4): the Assistant responds faster but may cut off users who pause mid-sentence. - Higher values (
0.6to1.0): the Assistant gives the user more time to continue speaking but adds latency before responding.
When set to null, the Assistant uses the integration's default endpointing value (see the Provider Support table above for provider-specific defaults). On an STS model the field has no effect at all, whatever its value.
Minimum value is 0, expressed in seconds. Typical range: 0.2 to 1.0.