Skip to main content
This page covers flow-level ASR configuration. For dynamic ASR biasing from Python functions, see conv.set_asr_biasing(). Setting up ASR biasing in the flow editor does not require code.
Automatic Speech Recognition (ASR) biasing helps the agent better understand the type of input it expects in each step. It nudges transcription toward expected patterns — like confirmation codes, personal names, or dates — instead of treating every utterance as free text. Enable it on any flow step that collects structured input.

​
When to use it

Structured input

The user is likely to say a booking code, date, or other structured value.

Frequent misrecognitions

You see transcription errors on names, codes, or numbers.

Unclear audio

The agent is guessing incorrectly from noisy or accented audio.

​
Bias options

Pick the bias that matches what the user is about to say — not the wording of the prompt. Custom keyword fields can also be added to bias transcription toward domain-specific vocabulary — product names, menu items, medical terms, and so on.
Steps with ASR biasing already applied are marked with the ear icon in the Flow Editor.

​
Configure it on a step

1

Open the step editor

In the Flow Editor, select the step where you want to bias transcription and open the step editor on the right.ASR biasing panel
2

Pick the input type

In the ASR biasing panel, toggle on the option that matches what the user is expected to say (see the table above).
3

Add custom keywords (optional)

Add domain-specific vocabulary as custom keywords — product names, menu items, dish names, medical terms. Bias applies on top of the selected input type.
4

Test with real voice input

Place a test call and confirm the transcription matches what was said. Iterate on the bias type and keywords if not.

​
Example: confirmation code capture

In the reservation confirmation flow example, the agent collects a booking code. Users say mixed letters and numbers, so Alphanumeric biasing is toggled on — this raises the likelihood that utterances like “X9C7G2” are transcribed accurately instead of being interpreted as ordinary words.
Combine with few-shot prompting when the input format is ambiguous — for example, when users might say a code as “X nine C seven G two” or “X9C7G2” interchangeably. Bias handles the acoustic model, few-shot prompting handles the LLM’s interpretation of the result.
Match the bias to the expected input, not the wording of the prompt. Biasing a “yes/no” prompt with Alphanumeric hurts recognition of “yes” and “no”.
Last modified on July 6, 2026