[Paper Review] Evaluation of a Multi-Resolution Dyadic Wavelet Transform Method for usable Speech Detection

Wajdi Ghezaiel, Amel Ben Slimane Rahmouni+1 more|arXiv (Cornell University)|Jan 2, 2013

Speech and Audio Processing3 citations

TL;DR

This paper proposes a multi-resolution dyadic wavelet transform (MRDWT) method to detect usable speech segments in co-channel speech mixtures, leveraging time-frequency analysis to isolate speech components. Evaluated on the TIMIT database, the method achieves 95.76% usable speech detection accuracy with 29.65% false alarms across diverse gender speaker mixtures.

ABSTRACT

Many applications of speech communication and speaker identification suffer from the problem of co-channel speech. This paper deals with a multi-resolution dyadic wavelet transform method for usable segments of co-channel speech detection that could be processed by a speaker identification system. Evaluation of this method is performed on TIMIT database referring to the Target to Interferer Ratio measure. Co-channel speech is constructed by mixing all possible gender speakers. Results do not show much difference for different mixtures. For the overall mixtures 95.76% of usable speech is correctly detected with false alarms of 29.65%.

Motivation & Objective

• Address the challenge of co-channel speech degradation in speaker identification systems.
• Improve the reliability of speech processing in overlapping speech environments.
• Develop a robust method to detect usable speech segments in mixed signals without prior speaker separation.
• Evaluate performance across diverse gender speaker combinations to ensure generalization.
• Assess detection accuracy and false alarm rates using the Target-to-Interferer Ratio (TIR) metric.

Proposed method

• Apply a multi-resolution dyadic wavelet transform (MRDWT) to decompose mixed speech signals into multiple frequency subbands.
• Use time-frequency representation to identify energy-dominant regions corresponding to active speech segments.
• Leverage the dyadic wavelet structure to achieve multirate analysis with logarithmic frequency resolution.
• Detect usable speech by thresholding wavelet coefficients based on energy concentration and temporal continuity.
• Utilize the TIMIT database for training and evaluation, simulating all possible gender speaker mixtures.
• Apply the Target-to-Interferer Ratio (TIR) as a performance metric to quantify detection quality.

Experimental results

Research questions

• RQ1 Can the MRDWT method effectively detect usable speech segments in co-channel speech mixtures?
• RQ2 How does detection performance vary across different gender speaker combinations in mixed signals?
• RQ3 What is the trade-off between usable speech detection accuracy and false alarm rate in overlapping speech?
• RQ4 To what extent does the multi-resolution nature of the dyadic wavelet transform enhance segment detection compared to single-resolution methods?
• RQ5 How does the method perform under realistic, diverse speaker mixture conditions?

Key findings

• The MRDWT method achieved a usable speech detection rate of 95.76% across all tested co-channel speech mixtures.
• The method exhibited a false alarm rate of 29.65%, indicating moderate sensitivity to non-speech or interferer-dominant segments.
• Performance remained consistent across different gender speaker combinations, with no significant variation in detection accuracy.
• The TIMIT database was successfully used to simulate realistic co-channel speech conditions for evaluation.
• The multi-resolution dyadic wavelet transform effectively captured time-frequency energy patterns associated with speech activity.
• The method demonstrated robustness in identifying usable speech segments despite the absence of prior speaker separation.