I performed this research with the Spoken Language Systems group at MIT CSAIL during Research Science Institute 2009.
Advances in automatic speech understanding bring a new paradigm of natural interaction with computers. The “Web-Accessible Multi-Modal Interface” (WAMI) system, developed at MIT’s CSAIL, provides a speech recognition service to a range of lightweight applications for Web browsers and cell phones.
In this research I addressed two limitations in the WAMI system. First, to improve performance, it required continual human intervention through expert tuning–an impractical endeavor for a large shared speech recognition system serving many applications. Second, WAMI was limited by its global set of models, suboptimal for its unrelated applications.
I developed a method to automatically adapt acoustic models and improve performance. At the same time, I extended the WAMI system to create separate models for each application. The system can now adapt to domain-specific features such as specific vocabularies, user demographics, and recording conditions. It also allows recognition domains to be defined based on any criteria, including gender, age group, or geographic location.
These improvements to WAMI bring it one step closer towards being an “organic,” automatically-learning system.
Tech Specs: The system automatically produces a training set from the utterances recognized with high confidence in the application context. I implemented this adaptive system and tested its performance using a large (100,000+ utterances) data set collected over one month from a voice-controlled game. The utterance error rate decreased by training with an adaptation set, and the trend suggests that accuracy will continue to improve with more usage.