Search

Recently, the focus of speech enhancement research has shifted from minimum mean-square error (MMSE) approaches, like the MMSE short-time spectral amplitude (MMSE-STSA) estimator, to state-of-the-art masking- and mapping-based deep learning approaches. We aim to bridge the gap between these two differing speech enhancement approaches. Deep learning methods for MMSE approaches are investigated in this work, with the objective of producing intelligible enhanced speech at a high quality. Since the speech enhancement performance of an MMSE approach improves with the accuracy of the used a priori signal-to-noise ratio (SNR) estimator, a residual long short-term memory (ResLSTM) network is utilised here to accurately estimate the a priori SNR. MMSE approaches utilising the ResLSTM a priori SNR estimator are evaluated using subjective and objective measures of speech quality and intelligibility. The tested conditions include real-world non-stationary and coloured noise sources at multiple SNR levels. MMSE approaches utilising the proposed a priori SNR estimator are able to achieve higher enhanced speech quality and intelligibility scores than recent masking- and mapping-based deep learning approaches. The results presented in this work show that the performance of an MMSE approach to speech enhancement significantly increases when utilising deep learning. Availability: The proposed a priori SNR estimator is available at: https://github.com/anicolson/DeepXi .

In this paper, we investigate the simultaneous modeling of multiple emotions in DNN-based expressive speech synthesis, and how to represent the emotional labels, such as emotional class and strength, for this task. Our goal is to answer two questions: First, what is the best way to annotate speech data with multiple emotions – should we use the labels that the speaker intended to express, or labels based on listener perception of the resulting speech signals? Second, how should the emotional information be represented as labels for supervised DNN training, e.g., should emotional class and emotional strength be factorized into separate inputs or not? We evaluate on a large-scale corpus of emotional speech from a professional voice actress, additionally annotated with perceived emotional labels from crowdsourced listeners. By comparing DNN-based speech synthesizers that utilize dierent emotional representations, we assess the impact of these representations and design decisions on human emotion recognition rates, perceived emotional strength, and subjective speech quality. Simultaneously, we also study which representations are most appropriate for controlling the emotional strength of synthetic speech.Keywords: Emotional speech synthesis, perception modeling, perceptual evaluation

This paper presents the real-time Simulink implementation of a noise adaptive speech processing pipeline for cochlear implants that was developed in a previous work. After providing an overview of each component or module in the pipeline, it is described how each module is implemented in Simulink so that the input audio frames are processed in real-time. This Simulink implementation allows the same code to be run on different hardware boards that are supported by Simulink. The performance of this implemented pipeline is evaluated in terms of six objective measures of speech quality. The results obtained indicate the effectiveness of suppressing noise in this speech processing pipeline when using the implemented automatic mechanism to identify the noise environment.

This paper focuses on Speech Quality Measurement for a satellite mobile communication system. In contrast to ground mobile communication systems, satellite speech quality measurement suffers from obvious jitter of long delays and severe satellite link losses. Auditory feature sensation and link loss measurement are most popular speech quality measurement models. However, long delay would cause the auditory feature sensation model with spreading of power spectrum, which furtherly diffuses the auditory spectrums to prevent human from hear correct responses. Nevertheless, measured error of link loss measurement would be enlarged gradually during the process of voice services. Therefore, a new speech quality measurement model based on the combination of auditory feature extraction from voice signal envelope and link loss from channel distortion is proposed. We analyze signal temporal envelope features and make power spectrum into auditory feature spectrum. The jitter of long delay and link loss are modeled as parameters to modify and compensate for the auditory feature spectrum, which also reduce the measurement distortion from transmission Doppler frequency offset to voice tone in satellite channel. A more understandable sensation estimation scores is proposed to present speech quality scores. The experimental results reveal that the new model reduces evaluation RMSE by 9.8% and is suited to the satellite mobile communications environment.

Voice transformation (VT) aims to change one or more aspects of a speech signal while preserving linguistic information. A subset of VT, Voice conversion (VC) specifically aims to change a source speakers speech in such a way that the generated output is perceived as a sentence uttered by a target speaker. Despite many years of research, VC systems still exhibit deficiencies in accurately mimicking a target speaker spectrally and prosodically, and simultaneously maintaining high speech quality. In this work we provide an overview of real-world applications, extensively study existing systems proposed in the literature, and discuss remaining challenges.

Performed subjective evaluations using subjects with various hearing impairments.Compared the perceptual effects of the three noise-reduction algorithms.No significant differences in the preference and intelligibility improvement.Significant differences in the quality improvement. Subjective evaluations of the perceptual effects of a specific noise-reduction (NR) algorithm for hearing-support devices have frequently been performed using normal-hearing (NH) participants because of financial, time, and ethical constraints. However, the perceptual effects of the same NR algorithm in a subjective evaluation may differ in accordance with the degree of hearing ability of the subjects. In this study, we performed subjective evaluations using 45 participants with different degrees of sensorineural hearing impairment to assess whether the perceptual effects of a certain NR algorithm in aspects of preference, quality, and intelligibility are affected by the degree of hearing ability. We recruited 15 NH subjects, 15 sensorineural hearing-impaired (SNHI) subjects with moderate hearing loss, and 15 SNHI subjects with moderately-severe hearing loss, and performed preference, intelligibility, and quality tests using three NR algorithms. Experimental results demonstrated that the perceptual effects of a certain NR algorithm were not significantly affected by the degree of hearing ability in regards to preference and improvement of speech intelligibility, but were significantly affected in regards to the improvement of speech quality.

Most noise reduction algorithms rely on obtaining reliable estimates of the SNR of each frequency bin. For that reason, much work has been done in analyzing the behavior and performance of SNR estimation algorithms in the context of improving speech quality and reducing speech distortions (e.g., musical noise). Comparatively little work has been reported, however, regarding the analysis and investigation of the effect of errors in SNR estimation on speech intelligibility. It is not known, for instance, whether it is the errors in SNR overestimation, errors in SNR underestimation, or both that are harmful to speech intelligibility. Errors in SNR estimation produce concomitant errors in the computation of the gain (suppression) function, and the impact of gain estimation errors on speech intelligibility is unclear. The present study assesses the effect of SNR estimation errors on gain function estimation via sensitivity analysis. Intelligibility listening studies were conducted to validate the sensitivity analysis. Results indicated that speech intelligibility is severely compromised when SNR and gain over-estimation errors are introduced in spectral components with negative SNR. A theoretical upper bound on the gain function is derived that can be used to constrain the values of the gain function so as to ensure that SNR overestimation errors are minimized. Speech enhancement algorithms that can limit the values of the gain function to fall within this upper bound can improve speech intelligibility.

This paper describes four experiments which have been carried out to evaluate the speech output component of the INSPIRE spoken dialogue system, providing speech control for di.erent devices located in a ‘‘smart’’ home environment. The aim is to quantify the impact of different factors on the quality of the system, when addressed either in the home or from a remote location (office, car). Factors analyzed in the experiments include the characteristics of the machine agent during the interaction (voice, personality), the physical characteristics of the usage environment (acoustic user interface, background noise, electrical transmission path), as well as task-related characteristics (listening-only vs. interaction situation, parallel tasks). The results show a significant impact of agent and environmental factors, but not of task factors. Potential reasons for this .nding are discussed. They serve as a basis for design decisions which have been taken for the final system.

The ultimate aim of the research reported on here is to develop an automatic testing system for Dutch pronunciation. In the experiment described in this paper automatic scores of telephone speech produced by native and non-native speakers of Dutch are compared with specific, i.e., temporal and segmental, and global pronunciation ratings assigned by three groups of experts: three phoneticians and two groups of three speech therapists. The goals of this experiment are to determinutee (1) whether specific expert ratings of pronunciation quality contribute to our understanding of the relation between human pronunciation scores and machine scores of speech quality, (2) whether different expert groups assign essentially different ratings, and (3) to what extent rater pronunciation scores can be predicted on the basis of automatic scores. The results show that collecting specific ratings along with overall ones leads to a better understanding of the relation between human and automatic pronunciation assessment. Furthermore, after normalization no considerable differences are observed between the ratings by the three expert groups. Finally, it appears that the speech quality scores produced by our speech recognizer can predict expert pronunciation ratings with a high degree of accuracy.

This paper addresses typical problems encountered with hands-free equipment in the context of GSM radiotelephony. We first summarise some important characteristics of the noise field in moving vehicles, and we also describe the acoustical echo phenomenon. We show that, in order to provide sufficiently high speech quality, these hands-free equipments should include noise reduction (NR) and acoustic echo control (AEC) devices. We then describe two possible structures combining noise reduction and acoustic echo control. As a conclusion, we raise the fact that the choice of a particular structure, among those proposed, is conditioned by the performance of the adaptation algorithm of the AEC solution.

Objective assessment techniques are computer-based algorithms to automatically assess speech quality withoutrequiring human listeners. When compared with listener-based tests, objective systems offer the potential advantages of low cost, quick turn-around and repeatable results. This paper briefly outlines four candidate objective techniques being considered by CCITT, and discusses their relative performance as determined during the CCITT 16 kbit/s LD-CELP tests. The results showed that none of the methods is sufficiently accurate or reliable to replace human listener panels in all applications. However, objective measuies may play an important role in future large-scale subjective evaluations as a means of obtaining coarse preliminary quality estimates for test design and validation.

COMSAT Laboratories acted as the processing host laboratory for Phases 1 and 2 of the CCITT Study Group XV 16 kbit/s speech coder evaluation program. The host laboratory was responsible for all preparations needed to allow the participating organizations to conduct subjective speech quality tests on the floating point hardware version of the 16 kbit/s LD-CELP coding algorithm. COMSAT was assisted by INTELSAT who provided funding, British Telecom and CPgD/Telebras (Brazilian Telecom Administration) who supplied personnel1 and by AT&T, Bellcore, National Institute of Science and Technology (NIST, a part of the US Department of Commerce), Nippon Telegraph & Telephone (NIT), and Philips Communication Industry AG (PKI) which supplied hardware/software. Additionally, the host laboratory worked with 12 laboratories (see Table 1) which provided source speech material in 9 different languages and conducted the subjective and objective tests. These efforts spanned the period between October 1989 and July 1990 for Phase I and from February to mid-April 1991 for Phase 2. The purpose of this paper is to describe the host laboratory functions that COMSAT performed. We also describe the Write Once and Read Many times (WORM) based data interchange system, defined by Bellcore and COMSAT, that was used to exchange data efficiently (a total of about 16,000 Mbytes) between participating laboratories. Finally, we describe the Common Analog Interface, built by NIT, which was used to ensure that all tested codecs received equal treatment.

The planned Pan-European cellular digital mobile radio (DMR) system envisaged for the early 1990s intends to use medium rate encoding of about 16 kbit/s. At least three phases of experimentation are planned, these being: (1) national pre-selection tests. (2) European selection test, and (3) the final characterisation and verification test with associated channel coding. The first two phases intend to use listening tests only where effects due to the extra processing delay of the codecs will not be tested. An initial maximum delay target of 65 ms was set. The final phase will use conversion tests do evaluate the effects of quality, delay and echo and will be complemented by listening tests which will also be used for codec characterisation. This paper describes subjective testing methodologies adopted to select suitable candidate codecs capable of being used in the proposed DMR system. Each administration participating in the tests had to convince the European Committee (Conference Europeenne des Administrations des Postes et Telecommunications, CEPT) that their candidate codec was better than or at least as good as the speech quality of companded FM analogue 900 MHz systems currently in use in Europe.