Transcriber (STT) configuration

The Transcriber section configures the speech-to-text step. It is organized into three sub-sections:

1. Provider / model / language — the primary transcriber.
2. Model Settings — provider-specific tuning.
3. Transcription Options — toggles for formatting and filtering.
4. Fallback Transcribers — what to try if the primary fails.

Provider, model, language

Pick the STT provider and model. Available providers include Deepgram, ElevenLabs, and Google.

Each provider exposes its supported models in the Model dropdown. For multilingual options, see Multilingual.

The API Key selector appears below the provider — use it to pick between the SimplAI-managed key and your own BYOK keys. See BYOK.

Model Settings

Keyterms

A comma-separated list of terms to improve recognition accuracy for domain-specific vocabulary (product names, jargon, acronyms).

• Supported only on Deepgram Nova-3 models. If you select a different model, the field is hidden.
• A token estimate (~N tokens) is shown below the field as you type.

Endpointing

Time in milliseconds of silence before speech is considered complete.

• Default: 25ms
• Range: 0 – 5000ms
• A value of 0 disables endpointing entirely.

Use longer endpointing values for users who pause mid-sentence; shorter values for snappier turn-taking.

Transcription Options

These toggles control how the transcriber formats and filters output. The set of available toggles depends on the provider — unsupported toggles are hidden, not greyed out.

Toggle	Default	Description
Use Numerals	Off	Convert spoken numbers from words to digits (e.g., "twenty five" → "25").
Profanity Filter	Off	Replace profanity in transcripts with asterisks.
Punctuate	On	Add punctuation to transcripts.
Smart Format	Off	Intelligent formatting for dates, currency, phone numbers, etc.
Filler Words	On	Include filler words (um, uh) in transcription. Recommended for better turn detection.
Interim Results	On	Provide preliminary results before final transcription is available.
Speaker Diarization	Off	Identify and label different speakers in the transcript.

Provider-specific visibility

• Deepgram — all toggles available. Keyterms only on Nova-3.
• ElevenLabs — only Interim Results is shown; the rest are handled internally by ElevenLabs.
• Google — Interim Results, Punctuate, and Speaker Diarization are shown; the rest are hidden.

Changing provider discards any values for toggles that are not supported by the new provider — no confirmation is shown, the fields simply disappear.

Fallback Transcribers

A collapsible section titled Fallback Transcribers lets you add backup providers for resilience.

• If no fallbacks are configured: No fallbacks configured. Add a fallback provider for improved reliability.
• Click + Add Transcriber Fallback to add a row with provider + model selectors and a trash icon to remove.
• You can add up to 3 fallback transcribers.
• Fallback order is top-to-bottom in the list — the first row is tried first.

How fallback runs at runtime

• A failure is an exception or a timeout (default 10s, configurable at the agent level).
• On failure, the next fallback is tried in order.
• Once a fallback succeeds, it is used for the rest of that turn. The next turn resets to the primary.
• When a fallback was used for a turn, the STT span in the trace carries a provider_fallback_used: true attribute — see Run history & traces.