Patterns

Patterns help recover short, structured alphanumeric tokens - order codes, reference numbers, licence plates - that are easy to mishear when spoken letter-by-letter and digit-by-digit. When you know the shape of a token that may appear in the audio, describe it with a subset of regular-expression syntax and the transcription is corrected to fit.

Patterns are available on the Prerecorded, Realtime, and Turns APIs.

Using Patterns

Pass the patterns query parameter as a comma-separated list of patterns. Each one describes a single token to recover.

?patterns=AMZ[0-9]{6}
?patterns=[A-Z]{2}[0-9]{2} [A-Z]{3}
?patterns=(INV)?[0-9]{4,5}
?patterns=AMZ[0-9]{6},ORD[0-9]{5}

Patterns use a subset of regular-expression syntax:

Syntax	Meaning	Example
[0-9]	a digit	[0-9]{6}
[A-Z]	a letter	[A-Z]{3}
{n}	repeat the preceding class exactly n times	[0-9]{6}
{m,n}	repeat the preceding class m to n times	[0-9]{4,6}
(...)?	an optional group - matched only when spoken	(INV)?[0-9]{5}
AMZ	literal letters/digits, e.g. a fixed prefix	AMZ[0-9]{6}
or -	a separator	[A-Z]{2} [0-9]{3}

Optional groups and ranges are resolved against the audio: an optional prefix is added only when it is actually spoken, and a {m,n} range adopts the spoken length. The more specific the pattern, the better - a fixed prefix or known structure (AMZ[0-9]{6}) is recovered more reliably than an open one.

A pattern matches tokens of at most 32 characters once fully expanded (the prefix and any digits or letters combined, separators excluded), and may describe at most 32 alternative shapes (each optional group and {m,n} range multiplies the alternatives). Syntax outside this subset - for example unbounded repeats like [0-9]+ - and patterns exceeding either limit are rejected with a 400 Bad Request.

Patterns cannot be combined with phrase biasing: don’t set patterns together with phrases or custom_model_id, or on requests from an API client with a custom model configured.

Examples

Token type	Example value	Pattern
Order code	AMZ374019	AMZ[0-9]{6}
Reference number	4429105538	[0-9]{10}
Licence plate	RU77 TVG	[A-Z]{2}[0-9]{2} [A-Z]{3}
Invoice (optional prefix)	INV4821 or 4821	(INV)?[0-9]{4}
Variable-length code	AMZ3740 … AMZ374019	AMZ[0-9]{4,6}

Only set patterns you expect

Patterns improve accuracy on audio that contains the token, but can degrade transcription of audio that does not. Set patterns only when the expected token is likely present in the request - don’t apply them globally.

See the API reference for Prerecorded, Realtime, and Turns.