To precisely model the time dependency of features is one of the important issues for speech recognition. Segmental unit input HMM with a dimensionality reduction method has been widely used to address this issue. Linear discriminant analysis (LDA) and heteroscedastic extensions, e.g., heteroscedastic linear discriminant analysis (HLDA) or heteroscedastic discriminant analysis (HDA), are popular approaches to reduce dimensionality. However, it is difficult to find one particular criterion suitable for any kind of data set in carrying out dimensionality reduction while preserving discriminative information. In this paper, we propose a new framework which we call power linear discriminant analysis (PLDA). PLDA can be used to describe various criteria including LDA, HLDA, and HDA with one control parameter. In addition, we provide an efficient selection method using a control parameter without training HMMs nor testing recognition performance on a development data set. Experimental results show that the PLDA is more effective than conventional methods for various data sets.

Acoustic feature transformation is widely used to reduce dimensionality and improve speech recognition performance. In this letter we focus on dimensionality reduction methods that minimize the average classification error. Unfortunately, minimization of the average classification error may cause considerable overlaps between distributions of some classes. To mitigate risks of considerable overlaps, we propose a dimensionality reduction method that minimizes the maximum classification error. We also propose two interpolated methods that can describe the average and maximum classification errors. Experimental results show that these proposed methods improve speech recognition performance.

In this study, toward the development of smartphone-based monitoring system for life logging, we collect over 1,400 hours of data by recording including both the outdoor and indoor daily activities of 19 subjects, under practical conditions with a smartphone and a small camera. We then construct a huge human activity database which consists of an environmental sound signal, triaxial acceleration signals and manually annotated activity tags. Using our constructed database, we evaluate the activity recognition performance of deep neural networks (DNNs), which have achieved great performance in various fields, and apply DNN-based adaptation techniques to improve the performance with only a small amount of subject-specific training data. We experimentally demonstrate that; 1) the use of multi-modal signal, including environmental sound and triaxial acceleration signals with a DNN is effective for the improvement of activity recognition performance, 2) the DNN can discriminate specified activities from a mixture of ambiguous activities, and 3) DNN-based adaptation methods are effective even if only a small amount of subject-specific training data is available.

We have developed an adaptation method which allows the customization of example-based dialog systems for individual users by applying “plus” and “minus” operations to the distributed representations obtained using the word2vec method. After retrieving user-related profile information from the Web, named entity extraction is applied to the retrieval results. Words with a high term frequency-inverse document frequency (TF-IDF) score are then adopted as user related words. Next, we calculate the similarity between the distrubuted representations of selected user-related words and nouns in the existing example phrases, using word2vec embedding. We then generate phrases adapted to the user by substituting user-related words for highly similar words in the original example phrases. Word2vec also has a special property which allows the arithmetic operations “plus” and “minus” to be applied to distributed word representations. By applying these operations to words used in the original phrases, we are able to determine which user-related words can be used to replace the original words. The user-related words are then substituted to create customized example phrases. We evaluated the naturalness of the generated phrases and found that the system could generate natural phrases.

In a distant-talking environment, the length of channel impulse response is longer than the short-term spectral analysis window. Conventional short-term spectrum based Cepstral Mean Normalization (CMN) is therefore, not effective under these conditions. In this paper, we propose a robust speech recognition method by combining a short-term spectrum based CMN with a long-term one. We assume that a static speech segment (such as a vowel, for example) affected by reverberation, can be modeled by a long-term cepstral analysis. Thus, the effect of long reverberation on a static speech segment may be compensated by the long-term spectrum based CMN. The cepstral distance of neighboring frames is used to discriminate the static speech segment (long-term spectrum) and the non-static speech segment (short-term spectrum). The cepstra of the static and non-static speech segments are normalized by the corresponding cepstral means. In a previous study, we proposed an environmentally robust speech recognition method based on Position-Dependent CMN (PDCMN) to compensate for channel distortion depending on speaker position, and which is more efficient than conventional CMN. In this paper, the concept of combining short-term and long-term spectrum based CMN is extended to PDCMN. We call this Variable Term spectrum based PDCMN (VT-PDCMN). Since PDCMN/VT-PDCMN cannot normalize speaker variations because a position-dependent cepstral mean contains the average speaker characteristics over all speakers, we also combine PDCMN/VT-PDCMN with conventional CMN in this study. We conducted the experiments based on our proposed method using limited vocabulary (100 words) distant-talking isolated word recognition in a real environment. The proposed method achieved a relative error reduction rate of 60.9% over the conventional short-term spectrum based CMN and 30.6% over the short-term spectrum based PDCMN.

A novel speech feature generation-based acoustic model training method for robust speaker-independent speech recognition is proposed. For decades, speaker adaptation methods have been widely used. All of these adaptation methods need adaptation data. However, our proposed method aims to create speaker-independent acoustic models that cover not only known but also unknown speakers. We achieve this by adopting inverse maximum likelihood linear regression (MLLR) transformation-based feature generation, and then we train our models using these features. First we obtain MLLR transformation matrices from a limited number of existing speakers. Then we extract the bases of the MLLR transformation matrices using PCA. The distribution of the weight parameters to express the transformation matrices for the existing speakers are estimated. Next, we construct pseudo-speaker transformations by sampling the weight parameters from the distribution, and apply the transformation to the normalized features of the existing speaker to generate the features of the pseudo-speakers. Finally, using these features, we train the acoustic models. Evaluation results show that the acoustic models trained using our proposed method are robust for unknown speakers.

We propose a blind dereverberation method based on spectral subtraction using a multi-channel least mean squares (MCLMS) algorithm for distant-talking speech recognition. In a distant-talking environment, the channel impulse response is longer than the short-term spectral analysis window. By treating the late reverberation as additive noise, a noise reduction technique based on spectral subtraction was proposed to estimate the power spectrum of the clean speech using power spectra of the distorted speech and the unknown impulse responses. To estimate the power spectra of the impulse responses, a variable step-size unconstrained MCLMS (VSS-UMCLMS) algorithm for identifying the impulse responses in a time domain is extended to a frequency domain. To reduce the effect of the estimation error of the channel impulse response, we normalize the early reverberation by cepstral mean normalization (CMN) instead of spectral subtraction using the estimated impulse response. Furthermore, our proposed method is combined with conventional delay-and-sum beamforming. We conducted recognition experiments on a distorted speech signal simulated by convolving multi-channel impulse responses with clean speech. The proposed method achieved a relative error reduction rate of 22.4% in relation to conventional CMN. By combining the proposed method with beamforming, a relative error reduction rate of 24.5% in relation to the conventional CMN with beamforming was achieved using only an isolated word (with duration of about 0.6 s) to estimate the spectrum of the impulse response.

To improve speech recognition performance, feature transformation based on discriminant analysis has been widely used to reduce the redundant dimensions of acoustic features. Linear discriminant analysis (LDA) and heteroscedastic discriminant analysis (HDA) are often used for this purpose, and a generalization method for LDA and HDA, called power LDA (PLDA), has been proposed. However, these methods may result in an unexpected dimensionality reduction for multimodal data. It is important to preserve the local structure of the data when reducing the dimensionality of multimodal data. In this paper we introduce two methods, locality-preserving HDA and locality-preserving PLDA, to reduce dimensionality of multimodal data appropriately. We also propose an approximate calculation scheme to calculate sub-optimal projections rapidly. Experimental results show that the locality-preserving methods yield better performance than the traditional ones in speech recognition.

Conventional features for Automatic Speech Recognition and Sound Event Recognition such as Mel-Frequency Cepstral Coefficients (MFCCs) have been shown to perform poorly in noisy conditions. We introduce an auditory feature based on the gammatone filterbank, the Selective Gammatone Envelope Feature (SGEF), for Robust Sound Event Recognition where channel selection and the filterbank envelope is used to reduce the effect of noise for specific noise environments. In the experiments with Hidden Markov Model (HMM) recognizers, we shall show that our feature outperforms MFCCs significantly in four different noisy environments at various signal-to-noise ratios.

To achieve high recognition performance for a wide variety of noise and for a wide range of signal-to-noise ratio, this paper presents methods for integration of four noise reduction algorithms: spectral subtraction with smoothing of time direction, temporal domain SVD-based speech enhancement, GMM-based speech estimation and KLT-based comb-filtering. In this paper, we proposed two types of combination methods of noise suppression algorithms: selection of front-end processor and combination of results from multiple recognition processes. Recognition results on the CENSREC-1 task showed the effectiveness of our proposed methods.

This paper introduces an evaluation framework for Japanese noisy speech recognition named AURORA-2J. Speech recognition systems must still be improved to be robust to noisy environments, but this improvement requires development of the standard evaluation corpus and assessment technologies. Recently, the Aurora 2, 3 and 4 corpora and their evaluation scenarios have had significant impact on noisy speech recognition research. The AURORA-2J is a Japanese connected digits corpus and its evaluation scripts are designed in the same way as Aurora 2 with the help of European Telecommunications Standards Institute (ETSI) AURORA group. This paper describes the data collection, baseline scripts, and its baseline performance. We also propose a new performance analysis method that considers differences in recognition performance among speakers. This method is based on the word accuracy per speaker, revealing the degree of the individual difference of the recognition performance. We also propose categorization of modifications, applied to the original HTK baseline system, which helps in comparing the systems and in recognizing technologies that improve the performance best within the same category.

Audio-Visual Speech Recognition (AVSR) is one of techniques to enhance robustness of speech recognizer in noisy or real environments. On the other hand, Deep Neural Networks (DNNs) have recently attracted a lot of attentions of researchers in the speech recognition field, because we can drastically improve recognition performance by using DNNs. There are two ways to employ DNN techniques for speech recognition: a hybrid approach and a tandem approach; in the hybrid approach an emission probability on each Hidden Markov Model (HMM) state is computed using a DNN, while in the tandem approach a DNN is composed into a feature extraction scheme. In this paper, we investigate and compare several DNN-based AVSR methods to mainly clarify how we should incorporate audio and visual modalities using DNNs. We carried out recognition experiments using a corpus CENSREC-1-AV, and we discuss the results to find out the best DNN-based AVSR modeling. Then it turns out that a tandem-based method using audio Deep Bottle-Neck Features (DBNFs) and visual ones with multi-stream HMMs is the most suitable, followed by a hybrid approach and another tandem scheme using audio-visual DBNFs.

Blind source separation is a technique that can separate sound sources without such information as source location, the number of sources, and the utterance content. Multi-channel source separation using many microphones separates signals with high accuracy, even if there are many sources. However, these methods have extremely high computational complexity, which must be reduced. In this paper, we propose a computational complexity reduction method for blind source separation based on frequency domain independent component analysis (FDICA) and examine temporal data that are effective for source separation. A frame with many sound sources is effective for FDICA source separation. We assume that a frame with a low kurtosis has many sound sources and preferentially select such frames. In our proposed method, we used the log power spectrum and the kurtosis of the magnitude distribution of the observed data as selection criteria and conducted source separation experiments using speech signals from twelve speakers. We evaluated the separation performances by the signal-to-interference ratio (SIR) improvement score. From our results, the SIR improvement score was 24.3dB when all the frames were used, and 23.3dB when the 300 frames selected by our criteria were used. These results clarified that our proposed selection criteria based on kurtosis and magnitude is effective. Furthermore, we significantly reduced the computational complexity because it is proportional to the number of selected frames.

To improve speech recognition performance, acoustic feature transformation based on discriminant analysis has been widely used. For the same purpose, discriminative training of HMMs has also been used. In this letter we investigate the effectiveness of these two techniques and their combination. We also investigate the robustness of matched and mismatched noise conditions between training and evaluation environments.