Noisy speech recognition failure diagnosis using Minimum Message Length decision trees

Current ASR technology lacks of effective failure diagnosis of ASR systems. Figures of merits such as WER are very useful, but don’t bring much insight into error patterns, error predictions or error analysis of ASR output. This paper explores an application of Minimum Message Length (MML) style decision trees for such a diagnosis, focusing on theoretical background and the failure diagnosis of noisy speech recognition. In addition, the paper focuses on failure diagnosis of noisy speech, covering several kinds of intrinsic speech variabilities as well. Results on added speech-shaped noise at different SNR, ranging from 25dB to −10dB, are presented.