Résumé: Abstract Speech recognition of tone languages requires detection of the tone in addition to detection of the consonants and vowels of a syllable. This paper compares three different methods based on the hidden Markov model (HMM) framework for recognition of tonal syllables in continuous speech. In joint detection, recognition is done by employing a HMM of connected tonal syllables, in which the pitch and its time derivative are included into the feature vector in addition to the phonetic features. In sequential detection, base syllables (syllables ignoring their tones) are recognized by using a HMM of connected base syllables only; the estimated syllable boundaries are then used for subsequent tone recognition in a separate HMM of tones. In linked detection, the recognition in the HMM of connected base syllables is modified to periodically take into account also tonal likelihood computed from a HMM of tones. Linked detection can provide performance that is comparable to the performance of joint detection, which is clearly superior to that of sequential detection. The computational complexity of linked detection is lower than that of joint detection for a large vocabulary task, where the number of states in the HMM of connected tonal syllables is substantially larger than that in the HMM of connected base syllables.

Langue: Anglais