A caller reads out their order code. The model drops one letter. The lookup fails, the wrong item ships, and your support queue grows by one. General-purpose speech recognition was built for natural language, not for the structured identifiers your transcripts actually depend on. That’s why we built Patterns, a feature on our API that biases the model toward those identifiers order codes, invoice numbers, license plates, SKUs, case references so they come out right the first time.
A general speech recognition model transcribes an order code the same way it transcribes a sentence: by predicting what’s most likely to come next in natural language. However, identifiers follow rules the model was never told about, which leads it to guess - dropping a letter, mis-formatting a licence plate, hearing “INB” instead of “INV”. When the result has to match a record exactly, every guess is a failed lookup, a wrong delivery, and another support ticket
You give us a sequence that describes the format you expect, and the model will bias towards it as it transcribes.
We measured the impact of Patterns on two relevant use cases, scoring word error rate on the identifier against the baseline model with no patterns applied.
Check out our documentation to learn more, and head over to our console or speak to our team to get started with Patterns today!