AssistantsConversation behavior

Voice formatting plan

Format LLM output for natural-sounding speech

Overview

Voice formatting automatically transforms raw text from your language model (LLM) into a format that sounds natural when spoken by a text-to-speech (TTS) provider. This process—called Voice Input Formatted—is enabled by default for all assistants.

Formatting helps with things like:

  • Expanding numbers and currency (e.g., $42.50 → “forty two dollars and fifty cents”)
  • Expanding abbreviations (e.g., ST → “STREET”)
  • Spacing out phone numbers (e.g., 123-456-7890 → “1 2 3 4 5 6 7 8 9 0”)

You can turn off formatting if you want the TTS to read the raw LLM output.

How voice input formatting works

When enabled, the formatter runs a series of transformations on your text, each handled by a specific function. Here’s the order and what each function does:

StepFunction NameDescriptionBeforeAfterDefaultPrecedence
1removeAngleBracketContentRemoves anything within <...>, except for <break>, <spell>, or double angle brackets << >>.Hello <tag> worldHello world✅-
2removeMarkdownSymbolsRemoves markdown symbols like _, `, and ~. Asterisks (*) are preserved in this step.**Wanted** to say *hi***Wanted** to say *hi*✅0
3removePhrasesInAsterisksRemoves text surrounded by single or double asterisks.**Wanted** to say *hi* to say❌0
4replaceNewLinesWithPeriodsConverts new lines (\n) to periods for smoother speech.Hello world\n to say\nWe have NASAHello world . to say . We have NASA✅0
5replaceColonsWithPeriodsReplaces : with . for better phrasing.price: $42.50price. $42.50✅0
6formatAcronymsConverts known acronyms to lowercase (e.g., NASA → nasa) or spaces out unknown all-caps words unless they contain vowels.NASA and .NETnasa and .net✅0
7formatDollarAmountsConverts currency amounts to spoken words.$42.50forty two dollars and fifty cents✅0
8formatEmailsReplaces @ with “at” and . with “dot” in emails.JOHN.DOE@example.COMJOHN dot DOE at example dot COM✅0
9formatDatesConverts date strings into spoken date format.2023 05 10Wednesday, May 10, 2023✅0
10formatTimesExpands or simplifies time expressions.14:0014✅0
11formatDistances, formatUnits, formatPercentages, formatPhoneNumbersConverts units, distances, percentages, and phone numbers into spoken words.5km, 43 lb, 50%, 123-456-78905 kilometers, forty three pounds, 50 percent, 1 2 3 4 5 6 7 8 9 0✅0
12formatNumbersFormats general numbers: years read as digits, large numbers spelled out, negative and decimal numbers clarified.-9, 2.5, 2023minus nine, two point five, 2023✅0
13removeAsterisksRemoves all asterisk characters from the text.**Bold** and *italic*Bold and italic✅1
14Applying ReplacementsApplies user-defined final replacements like expanding street abbreviations.320 ST 21 RD320 STREET 21 ROAD✅-

Customizing the formatting plan

You can control some aspects of formatting:

Enabled

Formatting is on by default. To disable, set:

1voice.chunkPlan.formatPlan.enabled = false

Number-to-digits cutoff

Controls when numbers are read as digits instead of words.

  • Default: 2025 (current year)
  • Example: With a cutoff of 2025, numbers above this are read as digits.
  • To spell out larger numbers, set the cutoff higher (e.g., 300000).

Replacements

Add exact or regex-based substitutions to customize output.

  • Example 1: Replace hello with hi: 
    1{ type: 'exact', key: 'hello', value: 'hi' }
  • Example 2: Replace words matching a pattern: 
    1{ type: 'regex', regex: '\b[a-zA-Z]{5}\b', value: 'hi' }

Currently, only replacements and the number-to-digits cutoff are customizable. Other options are not exposed.

Turning formatting off

To disable all formatting and use raw LLM output, set either of these to false:

1voice.chunkPlan.enabled = false
2// or
3voice.chunkPlan.formatPlan.enabled = false

Summary

  • Voice input formatting improves clarity and naturalness for TTS.
  • Each transformation step targets a specific pattern for better speech output.
  • You can customize or disable formatting as needed.