ANIQUE+: A new American national standard for non-intrusive estimation of narrowband speech quality

Abstract

Non-intrusive estimation of speech quality is a challenging problem in that the estimation has to be performed without the original reference speech signals. This paper presents a new American national standard (ANS) for nonintrusive estimation of narrowband speech quality. The proposed auditory non-intrusive quality estimation plus (ANIQUEϩ) model is a perceptual model simulating the functional roles of the human auditory system and employs improved modeling of quality estimation by a statistical learning paradigm. Experimental evaluation demonstrated that the performance of the ANIQUEϩ model is significantly superior to that of the current International Telecommunication Union-Telecommunication Standardization Sector (ITU-T) standard recommendation P.563 on 34 different subjective mean opinion score (MOS) databases. © 2007 Alcatel-Lucent.speech has to be available to estimate the quality of processed/degraded speech.In contrast to intrusive estimation, non-intrusive speech estimation from speech waveforms is a challenging problem in that the estimation of speech quality has to be performed with the speech signal under test only, without the reference speech signal information. Considering that reference speech signals are not presented together with the speech under test to human listeners in Absolute Category Rating (ACR) tests, non-intrusive estimation is more analogous to the situation of the most common type of subjective quality rating tests. Moreover, increasing demand for monitoring the speech quality of in-service networks focuses importance on the development of a non-intrusive speech quality estimation model. IntroductionThe need for automatic estimation of speech quality is becoming more prominent in modern telecommunication networks, as the reliable subjective assessment of the impact to the end-to-end speech quality of individual system components and combinations of them is extremely expensive in both time and cost, and most of the time, it is not feasible to explore all possible scenarios. Many objective models have been proposed and significant progress has been made in recent years. In 2001, the perceptual evaluation of speech quality (PESQ) was adopted as the P.862 standard recommendation by the International Telecommunication Union Telecommunication Standardization Sector (ITU-T) [12]. Most of models including P.862 are based on the intrusive paradigm, in which the reference Intrusive models cannot be used in this case since users' original reference speech signals uttered at the far end are not available in estimating the quality of near end speech.Difficulties in non-intrusive estimation originate mainly from our insufficient understanding as to how speech signals are processed in the human auditory systems for speech quality ratings. Despite these difficulties, there have been a series of works and progress in this area. Research on non-intrusive objective models had been unexplored in the literature until Liang and Kubichek proposed a vector quantization (VQ)-based non-intrusive model, in...

IEEE Signal Processing Magazine - November 2007

“…In a similar vein, reference-free models utilize prior knowledge of normal and/or abnormal behavior of speech signal features to estimate speech quality. To represent prior knowledge, existing reference-free mod els employ models of speech production [19], speech perception [46], speech signal feature likelihood [14], [43], or a combination thereof [48].…”

“…Two algorithms stood out during this competition; one became the ITU standard P.563 [32] and the other, ANIQUE+, became an American National Standard Institute (ANSI) standard [1], [46]. While these models have been shown to be reliable for many telecommunications scenarios, recent research has suggested that their quality prediction per formance is compromised for scenarios involving VQE algorithms (e.g., noise suppression [9] and dereverberation [16]) and wireless-VoIP tandem connections [12].…”

Reference-free automatic quality assessment of tracheoesophageal speech

“…In particular, we explore using the International Telecommunication Union ITU-T recommendation P.563 algorithm [7] and American National Standard Institute ANSI ANIQUE+ algorithm [8]. P.563 inputs a narrowband test speech signal and outputs an estimate of the subjective mean opinion score (MOS) of the signal.…”

“…These features and the estimated MOS serve as a potential set of candidate features for use in TE speech quality evaluation. To augment this candidate set we also include the MOS estimated by the ANIQUE+ algorithm [8]. ANIQUE+ estimates the MOS based on a perceptual model that mimics the human auditory system [8].…”

“…To augment this candidate set we also include the MOS estimated by the ANIQUE+ algorithm [8]. ANIQUE+ estimates the MOS based on a perceptual model that mimics the human auditory system [8]. In contrast to P.563, ANIQUE+ primarily relies on analysis of the test speech signal's modulation spectrum.…”