Convex weighting criteria for speaking rate estimation.

Area of Science:

• Speech Signal Processing
• Computational Linguistics
• Machine Learning for Audio Analysis

Background:

• Estimating speaking rate directly from speech waveforms is a challenging, long-standing problem.
• Existing methods often rely on phoneme detection or envelope heuristics, which can be unreliable.
• A robust and direct method for speaking rate estimation is needed.

Purpose of the Study:

• To develop a novel method for speaking rate estimation directly from speech waveforms.
• To avoid complex phoneme detection or heuristic-based approaches.
• To learn an optimal weighting function for direct application to time-frequency features.

Main Methods:

• The problem is framed as estimating a temporal density function from speech signals.
• An optimal weighting function is learned using convex cost functions.
• An adaptation strategy is proposed for speaker-specific customization with minimal training.
• The method is evaluated on TIMIT, dysarthric, and spontaneous speech corpora.

Main Results:

• The proposed method outperforms three competing methods on healthy and dysarthric speech.
• High correlation between estimated and ground truth speaking rates is observed for spontaneous speech.
• The approach effectively estimates speaking rate without explicit phoneme detection.

Conclusions:

• The novel weighting function approach provides an effective method for speaking rate estimation.
• The technique demonstrates robustness across different speech types, including disordered speech.
• This method offers a promising alternative to traditional speaking rate estimation techniques.