A speech mistranscription is an unexpected word or phrase produced when speech-to-text converts spoken language into text. It can distort a domain-specific request before the chatbot’s text classifier interprets it. section 4.4.3
When one word changes the request
An electric utility repeatedly received “presidential” where callers meant “residential.” Because the term affected call containment, the recurring error belonged in evaluation and training data while speech-model changes were pending. section 4.4.3
Handle suspected mistranscriptions by separating recognizable errors from unusable inputs:
- Train the speech models first when that capability is available, aiming for the most faithful utterance before text classification. section 4.4.3
- Add recurring, domain-significant mistranscriptions to the test set and eventually supplement the training data. section 4.4.3
- Set unintelligible transcriptions aside from blind testing and forward recognizable speech problems for further evaluation or to the appropriate team. section 4.4.2
