On duration and smoothing rules in a demisyllable-based isolated-word recognition system.

Area of Science:

• Speech Recognition
• Computational Linguistics
• Signal Processing

Background:

• Current isolated-word recognition systems often use demisyllable units to build word templates.
• Dynamic time warping (DTW) is a common algorithm for aligning speech patterns.
• Syllable duration variations between isolated and contextual speech pose a challenge.

Purpose of the Study:

• To investigate methods for improving isolated-word recognition accuracy.
• To evaluate the effectiveness of syllable duration preadjustment rules.
• To determine optimal parameter smoothing techniques for demisyllable units.

Main Methods:

• Constructing word reference templates by concatenating demisyllable units.
• Applying a dynamic time warping (DTW) algorithm for pattern alignment.
• Implementing and testing a syllable duration preadjustment rule for nonword-final stressed syllables.
• Investigating parameter smoothing at demisyllable boundaries with varying window durations and methods.

Main Results:

• A simple rule reducing rhyme demisyllable duration in stressed syllables achieved high recognition accuracy.
• Optimal smoothing window duration was found to be 60-90 ms.
• Failure to smooth reduced accuracy by only ~2% in a large test set.
• Linear and parabolic smoothing methods were equally effective.

Conclusions:

• Syllable duration preadjustment is crucial for effective isolated-word recognition.
• The demisyllable is a suitable basic unit for speech recognition systems.
• DTW remains a powerful tool, and parameter smoothing has a minor impact on accuracy.