Missing-data model of vowel identification

Area of Science:

• Auditory Neuroscience
• Speech Processing
• Acoustic Phonetics

Background:

• Vowel identity is linked to vocal tract transfer function formants.
• Voiced speech spectra are sampled at fundamental frequency (F0) harmonics, not directly reflecting formants.
• The auditory system's mechanism for deriving spectral envelopes from sampled vocalic waveforms remains unclear, especially with varying F0.

Purpose of the Study:

• To investigate how the auditory system identifies vowels despite the spectral distortion caused by fundamental frequency (F0) sampling.
• To propose a computational model for vowel identification that accounts for F0-dependent spectral undersampling and aliasing.

Main Methods:

• Modeled vowel identification as a pattern recognition problem with missing spectral data.
• Developed an F0-dependent weighting function to prioritize spectral regions near harmonics.
• Created frequency-domain (short-term spectra/tonotopic patterns) and time-domain (autocorrelation) versions of the model.

Main Results:

• The proposed model successfully accounts for the observed F0-independent nature of vowel identification.
• The model demonstrates how the auditory system can infer spectral envelopes from harmonic-sampled speech signals.
• The approach addresses spectral aliasing and distortion issues at high F0 values.

Conclusions:

• Vowel identification can be achieved through pattern recognition, effectively handling missing spectral information caused by F0 sampling.
• The model provides a plausible explanation for the robustness of vowel perception across different fundamental frequencies.
• This framework offers insights into auditory processing of complex harmonic spectra.