Title :
Sub-band speech recognition
Author :
Primor, David ; Furst-Yust, Miriam
Author_Institution :
Dept. of Electr. Eng.-Syst., Tel Aviv Univ., Israel
Abstract :
Automatic speech recognition (ASR) technology has become more accessible in the last decade. However, the performance of state of the art ASR systems is still far from being optimal in comparison to human performance. The robustness of common ASR systems is very limited. A possible improvement can be in the low-level acoustic-phonetic modeling. For example, improvement can be obtained by applying the recognition mechanism in parallel on nonoverlapping sub-bands. In order to show that such a mechanism can be beneficial, we have tested human ability to recognize speech embedded in a noisy background in non-overlapping sub-bands. The human performances were compared to a typical Hidden-Markov-Model (HMM) based ASR system (HTK). Consequently, we conclude that speech information exists in different and non-overlapping sub-bands with almost no significance to the central frequencies of the sub-bands. Using sub-band processes together with traditional processing, can obtain better automatic speech recognition.
Keywords :
acoustic signal processing; speech processing; speech recognition; ASR; automatic speech recognition; human performance; low-level acoustic-phonetic modeling; nonoverlapping sub-bands; robustness; sub-band speech recognition; Acoustic noise; Acoustic testing; Automatic speech recognition; Fatigue; Frequency; Humans; Robustness; Signal to noise ratio; Speech analysis; Speech recognition;
Conference_Titel :
Electrical and Electronics Engineers in Israel, 2002. The 22nd Convention of
Print_ISBN :
0-7803-7693-5
DOI :
10.1109/EEEI.2002.1178293