Advances in objective voice quality assessment

Abstract

The quality of transmitted voice is an important means of gauging performance of voice equipment and services. Such quality information has traditionally been provided by human listeners. Due to time and dollar costs of conducting listener tests, the need for automatic "objective" assessment techniques is growing rapidly. This paper reviews the status of objective voice quality technology, including four newer techniques being considered by CCITT. Recent performance tests based on a low bit-rate codec speech database are presented. Finally, the relative reliability and accuracy of objective and subjective techniques is discussed and compared.

Single-Ended Speech Quality Measurement Using Machine Learning Methods

“…Delta coefficients represent the local time derivatives (slope) of the cepstral sequence and are computed according to (5) Delta coefficients indicate the rate of change (speed) of spectral components; in our simulations is used. Double-delta coefficients are the second-order local time derivatives of the cepstral sequence and are computed according to (6) Double-delta coefficients indicate the acceleration of the spectral components; in our simulations, is used.…”

“…The speech databases are described in Table I. We separate 14 databases for training (databases [1][2][3][4][5][6][7][8][9][10][11][12][13][14] and the remaining six are used for testing (databases [15][16][17][18][19][20]. In addition, during training several algorithm parameters need to be optimized.…”

“…Earlier methods were implemented to assess the quality of waveform-preserving speech coders; representative measures include signal-to-noise ratio (SNR) and segmental SNR [4]. More sophisticated measures (e.g., [5]) were proposed once low-bit-rate speech coders, which may not preserve the original signal waveform, were introduced. More recently, quality measurement research has focused on algorithms that exploit models of human auditory perception.…”

Nonintrusive speech quality estimation using Gaussian mixture models

“…By adding a bias term to the model in [9], RMSE can be shown to be the sum of unexplained variance in the regression model, MOS estimation error due to limited number of listeners (affecting all algorithms equally), and the bias error between subjective MOS and objective MOS. The calculation of R does not take into consideration this bias error; therefore, unless the estimates are unbiased or all suffer from the same bias, RMSE is a more realistic measure of estimator performance.…”