Approach to Speech Processing

Artificial neural network based speech recognition by feature extraction analysisSpeech recognition by machine is a critical core technology for the "information" age. Existing machine recognition systems do not work the way human work. This is because automatic speech recognition (ASR) machines use spectral templates, while human work with partial recognition information across frequency, probably in the form of speech features that are local in frequency (e.g., formants). The forcing partial recognition errors in one frequency region do not affect the partial recognition at other frequencies (i.e., the partial recognition errors across frequency are independent). To extract the features spread across frequency requires frequency-local signal processing. Although a great deal has been learned, the fundamental process of speech production and speech perception, the goal of recognition of fluent speech remains elusive. The fundamental function of a speech recognition system is to identify trial speech utterances belonging to a given vocabulary with the highest possible degree, while rejecting those utterances that do not belong to the vocabulary.