Multiple task-domain acoustic models

Many speech recognition applications require the recognizer to perform at peak recognition accuracy across many different domains. Examples of different domains are general English, digits, names, alphabet, etc. Here we show a way to preserve the simplicity of a single acoustic model while providing domain specific recognition speed and accuracy. This is achieved by employing an extended phoneme set that keeps a subset of phonemes specifically for a particular domain, and a context dependency specification that allows cross-word, cross-domain phonetic context dependencies. Testing on a names recognition task going from a wrong domain (general English) model to a multiple domain model (general English, alphabet, names) the error rate is reduced by more than 50%. Domain-specific model trained only on the names data further reduces the error rate by more than 50%.