Speaker normalization of static and dynamic vowel spectral features

Area of Science:

• Speech processing
• Acoustic phonetics
• Machine learning for speech

Background:

• Speaker variability in acoustic features poses challenges for automatic speech recognition.
• Vowel spectral features, including formants and DCTCs, are crucial for distinguishing vowels.
• Existing methods often struggle with speaker-independent vowel identification.

Purpose of the Study:

• To evaluate two speaker normalization methods for vowel spectral features.
• To assess the impact of normalization on automatic vowel classification accuracy.
• To compare the effectiveness of different spectral parameters and feature types (static vs. dynamic).

Main Methods:

• Speaker normalization using multivariable linear transformation.
• Speaker normalization using polynomial frequency warping.
• Evaluation with formants and DCTCs (Discrete Cosine Transform Coefficients) as spectral parameters.
• Testing static and dynamic features, with and without fundamental frequency (F0).

Main Results:

• Both normalization methods consistently increased automatic vowel classification rates across all conditions.
• Linear transformation slightly outperformed polynomial warping.
• Normalized DCTCs trajectories achieved the highest classification rate (91%).
• Adding F0 improved non-normalized features but yielded diminishing returns with normalized features.

Conclusions:

• Speaker normalization is effective in improving automatic vowel recognition.
• Linear transformation is a slightly more effective normalization technique.
• Dynamic features combined with normalization show the most promise for robust vowel classification.