For voice agents (IVR) and voice applications, constrained grammar-based recognition delivers better accuracy, performance, and cost than traditional semantic-based speech-to-text (STT) AI engines. The engine processes grammars that follow the Speech Recognition Grammar Specification (SRGS) standard, which are usually written in an XML-based format called GrXML. The engine converts spoken audio (utterances) into text that follows the grammar's rules.

Constrained speech recognition works best for:

Recognizing constrained lists such as addresses, stock tickers, zip codes, and department names.
Recognizing alphanumeric strings like tracking numbers, account numbers, and confirmation codes that have positional constraints. For example, the first two characters of a member ID start with "AN", "FD", or "NT". Another example of positional constraints is vehicle identification number.
Alphanumeric or digit recognition with checksums or similar constraints. For example, credit card numbers with Luhn checksums.
Directed dialog applications requiring specific words or phrases.

Voice Agent Makers must perform the following steps:

Author GrXML grammar files within a rich text editor of choice that adhere to the Speech Recognition Grammar Specification standard.
After development, upload the grammar into an Azure Storage account.
In Microsoft Copilot Studio, create an external entity and use the URL of the grammar as the entity title.

During runtime, the voice agent loads the grammar, performs the recognition, and returns the required result.