Automatic detection of the second subglottal resonance and its application to speaker normalization

Area of Science:

• Speech processing
• Acoustic phonetics
• Speaker recognition

Background:

• Speaker normalization typically addresses vocal tract variations, causing spectral mismatch.
• Subglottal airways also influence speech spectra, with subglottal resonances showing promise for normalization.

Purpose of the Study:

• To develop a reliable algorithm for automatic estimation of the second subglottal resonance (Sg2).
• To investigate Sg2 frequency independence from speech content and language.
• To present a speaker normalization method utilizing Sg2.

Main Methods:

• Algorithm calibration using children's speech with accelerometer recordings for direct Sg2 measurement.
• Cross-language study with bilingual children (Spanish-English).
• Development and evaluation of an Sg2-based speaker normalization technique.

Main Results:

• The proposed algorithm reliably estimates Sg2 frequencies.
• Sg2 frequencies were found to be speaker-specific and largely independent of language and content.
• The Sg2 normalization method demonstrated superior performance compared to Vocal Tract Length Normalization (VTLN) for limited data and cross-language adaptation.

Conclusions:

• The second subglottal resonance (Sg2) is a robust feature for speaker normalization.
• The developed Sg2 estimation algorithm and normalization method offer an efficient and effective solution, particularly for limited data and cross-language scenarios.