Computational Modelling of Tone Perception Based on Direct Processing of f₀ Contours

Area of Science:

• Psychology
• Linguistics
• Computational Neuroscience

Background:

• Traditional speech perception models emphasize feature detection before phonetic unit recognition.
• The exact mechanism of feature extraction in continuous speech perception remains unclear.
• Auditory cues and articulatory gestures are considered potential features.

Purpose of the Study:

• To computationally model speech perception without intermediate featural representations.
• To investigate if phonetic categories can be recognized directly from continuous acoustic signals.
• To explore the operational mechanism of tone perception in Mandarin.

Main Methods:

• Utilized computational modeling with Support Vector Machine (SVM) and Self-organizing Map (SOM).
• Simulated Mandarin tone perception by directly processing fundamental frequency (f0) trajectories.
• Compared direct tone recognition with schemes involving tonal feature extraction.

Main Results:

• Direct tone recognition outperformed all tested feature extraction methods.
• Direct recognition required less computational power compared to feature-based approaches.
• The findings challenge the necessity of prior feature extraction in speech perception.

Conclusions:

• Prior extraction of features may not be the operational mechanism in speech perception.
• Direct processing of acoustic signals is a viable and efficient method for phonetic recognition.
• This study provides computational evidence against traditional feature-based speech perception theories.