Compensating for noise and mismatch in speaker verification systems using approximate Bayesian inference

This paper concerns effective speaker adaptation methods to solve the over-training problem in speaker verification, which frequently occurs when modeling a speaker with sparse training data. While various speaker adaptations have already been applied to speech recognition, these methods have not yet been formally considered in speaker verification. This paper proposes speaker adaptation methods using a combination of maximum a posteriori (MAP) and maximum likelihood linear regression (MLLR) adaptations, which are successfully used in speech recognition, and applies to speaker verification. Our aim is to remedy the small training data problem by investigating effective speaker adaptations for speaker modeling. Experimental results show that the speaker verification system using a weighted MAP and MLLR adaptation outperforms that of the conventional speaker models without adaptation by a factor of up to 5 times. From these results, we show that the speaker adaptation method achieves significantly better performance even when only small training data is available for speaker verification.

This study is focused on an unsupervised approach for detection of human scream vocalizations from continuous recordings in noisy acoustic environments. The proposed detection solution is based on compound segmentation, which employs weighted mean distance, T2-statistics and Bayesian Information Criteria for detection of screams. This solution also employs an unsupervised threshold optimized Combo-SAD for removal of non-vocal noisy segments in the preliminary stage. A total of five noisy environments were simulated for noise levels ranging from −20dB to +20dB for five different noisy environments. Performance of proposed system was compared using two alternative acoustic front-end features (i) Mel-frequency cepstral coefficients (MFCC) and (ii) perceptual minimum variance distortionless response (PMVDR). Evaluation results show that the new scream detection solution works well for clean, +20, +10 dB SNR levels, with performance declining as SNR decreases to −20dB across a number of the noise sources considered.

This paper presents the effect of mean normalization to various types of cepstral coefficients for robust speech recognition in noisy environments. Although the cepstral mean normalization (CMN) technique was originally designed to compensate channel distortion, it has also been proved that the CMN also improves recognition accuracy in additive noisy environment. However, no one has yet considered the interaction of CMN with spectral mapping functions required for extracting cepstral features. This paper investigates the impact of CMN to the speech recognition system depending on the types of spectral mapping function by mathematically analyzing the amount of spectral distortion between clean and noisy conditions. The analytic result is also confirmed by comparing the type of recognition error patterns in automatic speech recognition experiment with Aurora 2 database. Experimental results show that the performance improvement by adopting CMN becomes significant if the logarithmic function is replaced with the appropriate setting of fractional power mapping function. Especially, the deletion errors are dramatically reduced.

We present SocialSense , a collaborative smartphone based speaker and mood identification and reporting system that uses a user’s voice to detect and log his/her speaking and mood episodes. SocialSense collaboratively works with other phones that are running the app present in the vicinity to periodically send/receive speaking and mood vectors to/from other users present in a social interaction setting, thus keeping track of the global speaking episodes of all users with their mood. In addition, it utilizes a novel event-adaptive dynamic classification scheme for speaker identification which updates the speaker classification model every time one or more users enter or leave the scenario, ensuring a most updated classifier based on user presence. Evaluation of using dynamic classifiers shows that SocialSense improves speaker identification accuracy by 30% compared to traditional static speaker identification systems, and a 10% to 43% performance boost under various noisy environments. SocialSense also improves the mood classification accuracy by 4% to 20% compared to the baseline approaches. Energy consumption experiments show that its device daily lifetime is between 10-14 hours.

Speech recognition in noisy environments is one of the long-standing research themes but remains a very important challenge nowadays. Therefore, there is much research into all techniques and approaches to improve the performance of speech recognition systems, even in poor conditions. This paper presents a comparative study under various conditions based on two architectures (GMM-HMM and DNN-HMM), the Hybrid GMM-HMM models using the CMU Sphinx tools and the Hybrid DNN-HMM using the KALDI toolkit in noise environment. In this study, we compare the Hybrid GMM-HMM models and the Hybrid DNN-HMM models to evaluate the performance of the proposed system. The novelty of this paper is to test if the presented tools could be, with good accuracy, recognize the Arabic speech principally in noisy environment. In addition, we adopted the noisy training theory in this paper based on GMM-HMM and DNN-HMM model. We use the public Arabic Speech Corpus for Isolated Words (20 words), three noise levels, and three noise types. The implementation of our system consists of two phases: Features extraction using Mel-frequency Cepstral Coefficient (MFCC) and the classification phase will use separately the previous two models. In order to test the performance of these methods a simulation will presented for different SNR and for different district type of noise.

This paper describes advances in the use of HMM based technology for speaker independent continuous speech recognition, in noisy environment, under hands free interaction mode. For this purpose an array of four omnidirectional microphones is employed as the acquisition system. The processing of phase information in the cross-power spectrum provides the capability both of locating the talker position and of reconstructing an enhanced speech spectrum. Two enhancement techniques are described, that provide recognition improvement in the case of clean input speech as well as under different adverse conditions. The results refer to the use of a new multichannel corpus, collected in a real environment by a microphone array as well as a close-talk microphone.

All animals face the essential task of extracting biologically meaningful sensory information from the 'noisy' backdrop of their environments. Here, we examine mechanisms used by echolocating bats to localize objects, track small prey and communicate in complex and noisy acoustic environments. Bats actively control and coordinate both the emission and reception of sound stimuli through integrated sensory and motor mechanisms that have evolved together over tens of millions of years. We discuss how bats behave in different ecological scenarios, including detecting and discriminating target echoes from background objects, minimizing acoustic interference from competing conspecifics and overcoming insect noise. Bats tackle these problems by deploying a remarkable array of auditory behaviors, sometimes in combination with the use of other senses. Behavioral strategies such as ceasing sonar call production and active jamming of the signals of competitors provide further insight into the capabilities and limitations of echolocation. We relate these findings to the broader topic of how animals extract relevant sensory information in noisy environments. While bats have highly refined abilities for operating under noisy conditions, they face the same challenges encountered by many other species. We propose that the specialized sensory mechanisms identified in bats are likely to occur in analogous systems across the animal kingdom.

We propose a speaker-verification (SV) system to protect personal information stored on personal digital assistant (PDA) devices. PDAs have recently increased in capability and have increased in popularity, especially among highly mobile consumers. However, there still exist numerous obstacles to the practical use of SV with PDA devices because SV requires high computational complexity. To solve this problem, we propose a method to improve SV performance by applying the extended sequential probability ratio test (ESPRT) to the decision logic threshold. The proposed system is compatible with a PDA device's limited memory and low CPU speed.

Conventional speaker identification and speech recognition algorithms cannot deal with noisy and multiple speaker environments. For example, IBM via Voice has low recognition rates if dictation is done in a noisy environment. In order to achieve high performance in speaker identification and speech recognition, we propose an integrated approach that takes every facet of the process into account. Here we summarize some preliminary results from the application of this integrated approach to robust speaker identification and speech recognition. A real-time stand-alone software prototype has been developed to evaluate the effectiveness of the approach.

Automatic speaker recognition may achieve remarkable performance in matched training and test conditions. Conversely, results drop significantly in incompatible noisy conditions. Furthermore, feature extraction significantly affects performance. Mel-frequency Cepstral coefficients MFCCs are most commonly used in this field of study. The literature has reported that the conditions for training and testing are highly correlated. Taken together, these facts support strong recommendations for using MFCC features in similar environmental conditions (train/test) for speaker recognition. However, with noise and reverberation present, MFCC performance is not reliable. To address this, we propose a new feature 'Entrocy' for accurate and robust speaker recognition, which we mainly employ to support MFCC coefficients in noisy environments. Entrocy is the Fourier Transform of the Entropy, a measure of the fluctuation of the information in sound segments over time. Entrocy features are combined with MFCCs to generate a composite feature set which is tested using the Gaussian Mixture Model (GMM) speaker recognition method. The proposed method shows improved recognition accuracy over a range of signal-to- noise ratios.

Many power based features have been proposed in previous studies as alternative to the conventional feature, i.e. MFCC, for speech recognition. These features are of interest because they are empirically shown to be more robust than MFCC in noisy environments. Some studies argue that the compressions of power functions which are less sensitive than log for low energy spectra is one of the reasons while others agree that their performances are heavily affected by their normalization techniques. In this paper, the author focuses on investigating the effect of two normalization methods on power based features for the robustness of speech recognition in noisy environments. The paper specifically addresses features based on the q-log function, an example of power functions widely used in Tsallis statistics. The analysis suggests subtracting the mean may not be the best normalization technique due to its non-additive properties which is confirmed by the experimental results on Aurora-2. The results also confirm that the normalization may contribute more to the robustness of the features than the effect of the q-log compression.

Users of hearing aid devices encounter great difficulty in speech recognition in noisy environments, and show inferior performance in situations with interfering background noise or competing speakers as compared to the case of a quiet background. In such devices, a pre‐processing stage, which first removes the background interference, should greatly boost the speech recognition performance in noisy environments. In this work, the performance of a previously reported speech extraction algorithm is investigated for the purpose of improving the speech intelligibility of hearing aid users. Recently, we performed segregation of speech mixtures using our speech extraction algorithm and showed improvement in the speech recognition performance of normal hearing listeners. In this work, we investigate the performance of our algorithm for users of hearing aids on the same perceptual task of recognizing speech in the presence of either interfering speech or noise. We focus on tuning the segregation algorithm for hearing aid users, and emphasize on the preservation of the appropriate envelope and fine structure cues which will prove most helpful in the recognition of the target speech signal.

A text-dependent speaker verification system has been implemented for forensic application since 2008 when speech recording is allowed as legal evidence in Indonesian court. Due to the laborious procedures of the current system, an automatic text-independent speaker verification system is being developed in order to accelerate the forensic speaker verification process. The new forensic system is developed using a database spoken in Indonesian. The database is built on two scenarios, interview and conversation. This is done to simulate the real forensic condition for speaker verification in Indonesia. The system adopts the very famous Gaussian Mixture Model (GMM) and Universal Background Model (UBM) framework in speaker verification field and using Mel-Frequency Cepstral Coefficients (MFCC) as the speech feature. Along with the application of zero-normalization, the new forensic speaker verification system has successfully reached an Equal Error Rate (EER) of 4.66 % that is a better performance than the previous developed system of Indonesian forensic speaker verification.

At present, most speaker recognition algorithms are performed in a clean environment, and the effect is poor in a noisy environment. In order to improve the accuracy of speaker recognition in a noisy environment, a new feature extraction method, wavelet packet & Gammatone (WPGT) based model, is proposed. In this model, the wavelet packet is used to decompose high-frequency and low-frequency signals and the Gammatone filter bank simulates the human auditory system to process non-linear signals so that more complete speaker voice features are extracted, and finally, the convolutional neural network is used to train the features and complete speaker recognition. Based on the open source speech data sets and the noise fusion data sets, the proposed method is compared with the commonly used voiceprint feature extraction methods MFCC and Gammatone. The experimental results show that, in a noisy environment, WPGT has better anti-noise ability than MFCC and Gammatone. Compared with MFCC and Gammatone, the accuracy of WPGT is improved by 10.63% and 16.91%, respectively.

The mismatch between the training conditions and the test conditions severely degrades the performance of speaker verification. Aiming at solving this problem, this paper presents a method which uses a combination of improved nonnegative matrix factorization (IMNMF) and mel frequency cepstral coefficients (MFCC) with feature warping for improving identity-vector (i-vector) speaker verification performance in noisy environment. Unlike the traditional nonnegative matrix factorization (NMF), IMNMF uses extra free basis vectors to capture the features which are not included in training data, and linear constraints on dictionary atoms. Feature warping is used to remove channel noises. Therefore, the proposed method can reduce distortion of reconstructed speech while enhancing the recovered speech quality. The performance of i-vector speaker verification is evaluated by using the short utterance database and the NOISEX-92 database. The experiment results indicate that the score level fusion of feature-warped IMNMF-MFCC and feature-warped MFCC is superior to the baseline system at the equal error rate under the majority of signal-to-noise ratios (SNRs).