Fine-grained statistical structure of speech.

Area of Science:

• Speech processing
• Auditory perception
• Signal analysis

Background:

• Speech signals exhibit statistical regularities exploitable for efficient coding.
• Time-frequency representations with power-law frequency selectivity capture speech structure.
• Variations in the power-law exponent adapt to phonetic categories.

Purpose of the Study:

• Investigate variations of the power-law exponent in speech at a finer level.
• Relate exponent values to key acoustic properties of speech.
• Explore efficient coding strategies in speech perception.

Main Methods:

• Parametric approach using Gabor dictionaries for signal decomposition.
• Estimation based on sparsity of decompositions.
• Analysis at phoneme and intra-phonemic levels.

Main Results:

• Exponent variations correlate with phonetic categories and acoustic properties.
• Plosive release bursts lower the exponent; vowels' exponent relates to formant bandwidths.
• Efficient coding reduces frequency selectivity with sound intensity, matching cochlear nonlinearities.

Conclusions:

• Speech's fine-grained statistical structure is interpretable via exponent variations.
• Acoustic properties like release bursts and formant bandwidths influence the exponent.
• Efficient coding strategies in speech perception align with biological auditory processing.