Tone Recognition With Fractionized Models and Outlined Features
Ye Tian, Jian-Lai Zhou, Min Chu, Eric ChangICASSP 2004
Hsiao-Tsung Hung
Department of Computer Science and Information EngineeringNational Taiwan Normal University
Outline
• Introduction• Features– Detailed features– Outlined features– Experiments and analysis
• Tone Modeling– Experiments and analysis
• Conclusions
Introduction
• 2 questions1. Is the detailed information of F0 curve useful
for tone discrimination in continuous speech?
2. Are phoneme-independent tone models sufficient for continuous speech recognition?
Detailed features
• Detailed features: Using the entire F0 curve.• Observation vector is • If the phoneme has totally N frames, the
number of total parameters used for tone recognition is 2*N.
Outlined features
• To reduce the number of parameters and improve the robustness.
1. Curve fitting features2. Subsection Outlined features
Curve fitting features
• First-order
• Second-order
Subsection Outlined features
• The F0 curve of the entire phoneme is divided into several subsections and each subsection is represented by certain parameters.
• Extract parameters for each subsection– 1.subsection slop and intercept– 2.subsection and (Assume that time frames belong to the subsection k.)
Y
X
X={0,1,…, //frame
Y=
F0
Subsection Outlined features
1.subsection slop and intercept
Subsection Outlined features
2.subsection and
Experiments and analysis
1.Main value and direction are the most important characteristics.
2.Detailed information is useless for tone discrimination.
Tone Modeling
1. One-tone-one-model tone models(5)2. Monophone-dependent tone models(54)
The same tone in different tonal phonemes is different modeled.
3. Triphone-dependent tone models(12824)
Experiments and analysis
• Feature vector :
Conclusions
• Using fractionized models and outlined features for tone recognition.
• Outlined features can reduce the interference caused by co-articulation effect, syllable stress, and sentence intonation.
