Objective speech quality measurement using statistical data mining

Authors:
Wei Zha;Wai-Yip Chan
Affiliations:
Power, Acquisition and Telemetry Group, Schlumberger Technology Corporation, Sugar Land, TX;Department of Electrical & Computer Engineering, Queen's University, Kingston, ON, Canada
Venue:
EURASIP Journal on Applied Signal Processing
Year:
2005

Citing 6
Cited 0

Speech Coding and Synthesis

Speech Coding and Synthesis
Digital Coding of Waveforms: Principles and Applications to Speech and Video

Digital Coding of Waveforms: Principles and Applications to Speech and Video
Psychoacoustics: Facts and Models

Psychoacoustics: Facts and Models
Nonlinear prediction of mobile radio channels: measurements and MARS model designs

ICASSP '99 Proceedings of the Acoustics, Speech, and Signal Processing, 1999. on 1999 IEEE International Conference - Volume 05
Perceptual evaluation of speech quality (PESQ)-a new method for speech quality assessment of telephone networks and codecs

ICASSP '01 Proceedings of the Acoustics, Speech, and Signal Processing, 200. on IEEE International Conference - Volume 02
Bayes risk weighted vector quantization with posterior estimation for image compression and classification

IEEE Transactions on Image Processing

Quantified Score

Hi-index	0.00

Visualization

Abstract

Measuring speech quality by machines overcomes two major drawbacks of subjective listening tests, their low speed and high cost. Real-time, accurate, and economical objective measurement of speech quality opens up a wide range of applications that cannot be supported with subjective listening tests. In this paper, we propose a statistical data mining approach to design objective speech quality measurement algorithms. A large pool of perceptual distortion features is extracted from the speech signal. We examine using classification and regression trees (CART) and multivariate adaptive regression splines (MARS), separately and jointly, to select the most salient features from the pool, and to construct good estimators of subjective listening quality based on the selected features. We show designs that use perceptually significant features and outperform the state-of-the-art objective measurement algorithm. The designed algorithms are computationally simple, making them suitable for real-time implementation. The proposed design method is scalable with the amount of learning data; thus, performance can be improved with more offline or online training.