Hidden Markov model for Mandarin lexical tone recognition

Tone nucleus modeling for Chinese lexical tone recognition

“…Hidden Markov models (HMMs) were found to be an efficient method for pitch tone recognition in a number of previous studies (Ljolje and Fallside, 1987;Yang et al, 1988;Hirose and Hu, 1995;Wang et al, 1997). We also used HMMs as the tonal acoustic models in the tone recognition system.…”

“…Previous studies have shown that it is easy to recognize lexical tones in isolated speech (Yang et al, 1988;Le et al, 1993), but rather difficult to recognize them from F0 contours of continuous speech (Wang et al, 1997;Liu et al, 1999). This different performances can be ascribed to the reason that the lexical tones show consistent tonal F0 patterns when uttered in isolation, but show complex variations in continuous speech.…”

Automatic speech recognition and speech variability: A review

“…Tone can be modeled separately through specific HMMs (Yang et al, 1988) or decision trees (Wong and Siu, 2004), or the pitch parameter can be included in the feature vector (Chen et al, 1997), or both information streams (acoustic features and tonal features) can be handled directly by the decoder, possibly with different optimized weights (Shi et al, 2002). Various coding and normalization schemes of the pitch parameter are generally applied to make it less speaker dependent; the derivative of the pitch is the most useful feature (Liu et al, 1998), and pitch tracking and voicing are investigated in Huank and Seide (2000).…”

Tone nucleus-based multi-level robust acoustic tonal modeling of sentential F0 variations for Chinese continuous speech tone recognition

“…It was found that high performance was easy to achieve in the tone recognition of isolated syllables or short words (2-4 syllables) (Yang et al, 1988;Wang et al, 1990), but that continuous speech presented difficulties that resulted in a much lower performance (Wang et al, 1997;Cao et al, 2000).…”