Speech recognition and caller input

Choose how the agent hears callers: language, recognition engine, noise removal, medical vocabulary, keypad input and what personal data is redacted.

These settings control what the agent hears. They sit on the agent alongside the tuning options, and every field is in the API reference.

Language and recognition

Setting	Values	What it does
`language`	A language code such as `en` or `es`	The language the agent listens for.
`stt_model`	`whisper`, `conformer`, `deepgram`	The speech recognition engine. `conformer` is English only: choose another language with it and the agent uses `whisper` instead.
`stt_mode`	`fast`, `accurate`	Speed or accuracy. It applies only to the `deepgram` engine. `accurate` is used for English with general vocabulary when it is available for your account. Otherwise the agent stays on `fast` rather than failing the call.
`vocab_specialization`	`general`, `medical`	`medical` tunes recognition for clinical terms.
`boosted_keywords`	List of words	Words to favour, such as product and people's names.
`enable_transcription_formatting`	true / false	Adds punctuation and number formatting to transcripts. Off by default because the extra step can slow the reply.

If you do not set stt_model, a new agent uses Dialgem's default for your account.

Noise removal

denoising_mode cleans the caller's audio before it is transcribed.

Value	Effect
`off`	No processing. The default.
`noise-cancellation`	Removes steady background noise.
`noise-and-background-speech-cancellation`	Also suppresses other people talking behind the caller.

denoising_strength (0 to 1, default 0.5) sets how hard the last mode suppresses. Stronger settings remove more noise but can clip the caller's own voice, so raise it gradually and listen to a test call.

Knowing when the caller has finished

adaptive_response_wait makes the agent wait a little longer when the caller sounds unfinished: a trailing "and", a comma, or half of a number read out.
backchannel_frequency (0 to 1, default 0.8) sets how often the agent says "mm-hm" while the caller talks. It only matters when enable_backchannel is on.

Keypad input

With allow_user_dtmf on, the digits a caller presses are collected and given to the agent as one caller turn, such as (pressed on keypad) 1234. This suits account numbers, extensions and PINs that are awkward to say aloud.

user_dtmf_options decides when an entry is complete, whichever happens first:

Option	Meaning
`digit_limit`	Complete after this many digits (1 to 50). Unset means no limit.
`termination_key`	Complete when the caller presses this key (`0`-`9`, `#` or `*`). The key itself is not included.
`timeout_ms`	Complete after this long with no further digit (500 to 15000, default 2500).

Keypad input works on calls carried by Twilio. Calls over a SIP trunk do not pass digits to the agent reliably yet, so test your own trunk before relying on it. To make the agent press keys itself, use the press digit function.

Redacting personal data

When pii_redaction_enabled is on, personal data is removed from stored transcripts. pii_categories chooses which kinds:

email
credit_card
ssn
phone_number

The default is all four. Redaction is applied to the transcript and event log that are stored for the call (and any digits pressed on the keypad are removed from the log). It also means a supervisor cannot listen in or take over on that agent's calls, because live audio cannot be redacted. See Data retention for how long stored data is kept.