Stress and vowel duration effects on syllable recognition

Area of Science:

• Acoustic Phonetics
• Speech Recognition Technology
• Computational Linguistics

Background:

• Continuous speech recognition systems face challenges adapting to acoustic variations, particularly stress.
• Speaker-dependent recognition requires robust methods to handle variations in pronunciation.
• Understanding the acoustic differences between stressed and unstressed speech is crucial for improving recognition accuracy.

Purpose of the Study:

• To investigate the acoustic differences between stressed and unstressed syllables in continuous speech.
• To determine if separate templates are needed for stressed and unstressed syllables in speech recognition systems.
• To compare human and algorithmic recognition performance on stressed and unstressed syllables.

Main Methods:

• A speaker-dependent study using excised syllables from continuous speech.
• Transformation of syllables into cepstral coefficient arrays at two precision levels.
• Formation of four template dictionaries varying in size and stress composition using time-warping.
• Collection of recognition performance data from human listeners and a computer algorithm.

Main Results:

• Stressed and unstressed versions of the same syllable are often acoustically distinct, justifying separate templates.
• Unstressed syllables show acoustic variance comparable to stressed syllables.
• Long vowels reduce cross-vowel errors but can impair consonantal discrimination by algorithms.
• Listener consonant error patterns differ significantly from algorithmic patterns based on vowel length.

Conclusions:

• Separate acoustic templates for stressed and unstressed syllables are beneficial for speech recognition.
• Acoustic variance is similar between stressed and unstressed syllables.
• Vowel length impacts both error types and algorithmic discrimination capabilities.
• Discrepancies exist between human and algorithmic error patterns, highlighting areas for further research.