A Pragmatic View of the Application of HMM2 for ASR

This report investigates the HMM2 approach recently introduced in the framework of automatic speech recognition. HMM2 can be seen as a mixture of HMMs, where a conventional primary HMM (processing a time series of speech data) is supported on a lower level by a secondary HMM, working along the frequency dimension of a temporal segment of speech. The application of HMM2 to the speech signal is motivated by numerous potential advantages. However, speech recognition results did not show the expected performance improvements. In this paper, the HMM2 approach is pragmatically analyzed and evaluated on speech data, revealing some problems and suggesting potential solutions.