Speech recognition of mandarin monosyllables

摘要: The nonlinear dynamic characteristics of expansion and contraction the sequential time-varying features syllable pronunciations greatly complicate tasks automatic speech recognition. Each is represented by a sequence vectors linear predict coding cepstra (LPCC). Even if same speaker utters syllable, duration stable parts LPCC changes every time. Therefore, contracted such that compressed waveform has length. We propose five different simple techniques to contract vectors. A simplified Bayes decision rule with weighted variance used classify 408 speaker-dependent mandarin syllables. For syllables, recognition rate 94.36% as compared 79.78% obtained using hidden Markov models (HMM). 98.16% achieved within top 3 candidates. proposed in this paper represent syllables are easy be extracted. computation for feature extraction classification much faster than HMM or any other known techniques.