Search

This paper proposes a new speech coding method pitch synchronous innovation code excited linear predictor (PSI-CELP). This method is based on CELP but adds pitch synchronous innovation. This results in even random code vectors being adaptively converted to have pitch periodicity for voiced frames. This scheme can improve the synthesized speech quality of voiced frames in the low bit-rate CELP without increasing either computational complexity or bit rate. In addition to pitch synchronous innovation, this paper also proposes four other methods in which the quality of synthesized speech is effectively improved or the computational complexity is reduced for low bit-rate CELP: (1) the perceptual weighting filter is an MA-type and uses nonquantized LPC parameters; (2) the fixed codebook is prepared as part of the adaptive codebook for nonperiodic frames; (3) the random codebook consists of two-channel conjugate sub-codebooks; and (4) codebook searches are made using delayed decision. In a speech coding system with pitch synchronization of random code vectors, improvements in ratio SNR, segmental SNR and opinion equivalent Q value are respectively 1.6 dB. 1.0 dB. and 0.5 dB. These basic methods were adopted in the standard code for Japanese half-rate digital mobile telephone service. [ABSTRACT FROM AUTHOR]

This paper discusses parameter estimation for a continuous density HMM (hidden Markov model) method of visual speech recognition. Past studies of visual speech recognition can be broadly divided into two approaches, the image-based method and the model-based method. The image-based method is a method in which some preprocessing, such as subsampling and principal component analysis, is applied to the pixel values of the original image, and the result is used as the feature vector. In this approach, the position and size of the lips and the illumination conditions have a direct effect on the recognition rate. Thus, the normalization of these factors is the basic technique. The ordinary approach in the conventional normalization is to provide a criterion independently of the HMM, and to apply normalization before learning. In this paper, normalization by the ML (maximum likelihood) criterion is considered. Normalized training is proposed in which the normalization processes for elements such as the position, size, inclination, mean brightness, and contrast of the lips are integrated with the training of the model. The proposed method is formulated on the basis of an EM (expectation maximization) algorithm in which monotonically increasing behavior of the likelihood of the training data is guaranteed, by iteration of normalized training. The effectiveness of the proposed method is demonstrated in a word recognition experiment using the M2VTS database. © 2006 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 89(11): 40–50, 2006; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.20281 [ABSTRACT FROM AUTHOR]

This paper proposes a spectral estimation method of speech based on mel-generalized cepstral representation. The spectral estimation based on generalized cepstral representation is a method that unifies the cepstral method and the linear prediction method. By this method, the spectral model can be varied continuously from the all-pole model to the exponential model. This paper also discusses the following: the method of spectral estimation, where the mel-generalized cepstnim is used instead of the generalized cepstrum; the uniqueness of the solution for the minimization problem; the solution method; and the convergence of the solution. The proposed method has the same features as those of the method based on the generalized cepstrum where the spectral model can be varied continuously and the stability of the obtained transfer function is guaranteed. Another feature is that characteristics similar to the human auditory sensation, i.e., the high resolution in the low-frequency range and the coarse resolution in the high-frequency range, are obtained since the mel-generalized cepstrum is used as the parameter, which is the generalized cepstrum defined on the mel-frequency axis. Finally, the features and the effectiveness of the proposed method is demonstrated by analysis examples of the synthesized signals and natural speech. [ABSTRACT FROM AUTHOR]

Pattern recognition methods in which multiple templates are prepared for each category and which recognize input patterns using those templates usually partition training patterns into clusters and generate templates as centroids of each cluster. However, most clustering algorithms proposed in the literature aim at minimizing the square error distortion measure when each cluster is represented by Its centroid. Therefore the result obtained is not necessarily optimal iron the viewpoint of discrimination. This paper describes a new clustering algorithm which is a method of obtaining an optimum solution to a template generation problem. The simulated annealing is used for searching for a cluster partition that minimizes the "degree" of recognition error when all training patterns are recognized by using centroids as templates. This paper also describes a vowel template generation problem in speech recognition as an example of pattern discrimination problems. The new algorithm yielded better experimental results than the LBC algorithm and the LVQ algorithm. [ABSTRACT FROM AUTHOR]

One of the problems with the hidden Markov model (HMM) in performing speech recognition is that the local transition information of the feature vectors is not incorporated into the mechanism of the model and the model is not constrained by transitions of the feature vectors. Thus, the output probability distribution never changes during recognition. Furthermore, all transitions between the vectors that have high probabilities are allowed even if those transitions did not appear in the training data. This paper proposes a bigram-constrained HMM that uses correlations between two frames to constrain the feature distributions of a speaker-independent HMM to the region most appropriate for the speaker. Since the output probability of the bigram-constrained HMM is a conditional probability restricted by the feature vector of the previous frame, the output probability changes dynamically at each frame depending on the feature vector of the previous frame. Constraining the feature distribution makes it possible to reduce the overlapping of feature distributions between different phonemes which improves recognition performance. Previously, we proposed the discrete bigram-constrained HMM which is based on the combination of a discrete speaker-independent HMM and the VQ-code bigram. We showed that it performed better than conventional speaker-independent HMMs. In this paper, the strategy is extended to the tied-mixture bigram-constrained HMM and the continuous bigram-constrained HMM to obtain better recognition performance. These three types of HMMs are formulated and evaluated by phoneme recognition in continuous speech. [ABSTRACT FROM AUTHOR]

The realization of an anthropomorphic agent which looks like a real human is an important research topic for the broadening of the range of human-to-human communications through the use of a computer. We have proposed a technique for synthesizing natural talking-face animation that permits such communications. How to evaluate the performance of talking-face animation, however, has remained an outstanding issue. The performance of talking-face animation is determined in three parameters: (1) Does it reproduce human talking to an extent that permits lipreading? (2) Does it appear visually natural? (3) Is it accurately synchronized with voice? In this paper, we first presented talking-face animation along with the voice to subjects and conducted experiments on how well the subjects heard the contents of the spoken words to examine Parameter (1). In the next step, with regard to Parameter (2), the visual naturalness of the talking-face animation and the smoothness of the motion of the talking mouth were evaluated on a scale of 5 points. Lastly, with regard to Parameter (3), talking-face animation in which the synchronization of the animation with sound was off by a fixed interval was shown to subjects to investigate the subjective perception of the synchronization gap, and the extent of the resulting strange feeling was evaluated on a scale of 5 points. In addition, the effect of the synchronization gap between voice and talking-face animation on the manner in which the spoken words are understood was also evaluated. Through these evaluation experiments, the quality of synthetic talking-face animation proposed by the authors was evaluated, and we studied naturally-appearing synchronization between synthetic talking-face animation and voice. © 2006 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 89(5): 39–52, 2006; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.20180 [ABSTRACT FROM AUTHOR]

This paper considers the continuous distribution HMM and discusses speed improvement by parallel processing and VLSI design, aiming at the realization of a high-speed speech recognition LSI. The continuous distribution HMM has high recognition performance. But a large amount of computation is required for the calculation of the output probability, and speed improvement is highly desirable. Regarding speed improvement of the output probability calculation, this paper proposes blockwise parallel processing and an architecture for performance. The minimum optimal bit width is investigated for recognition by fixed-point operations. Based on the proposed architecture, VLSI design for an HMM is performed and the result is evaluated. It is shown that the proposed architecture has sufficient power to construct a recognition system that can respond in real time, requiring 3 ms for the recognition processing of 100 words. © 2004 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 87(5): 12–23, 2004; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.10099 [ABSTRACT FROM AUTHOR]

The digitization and restoration of images that have been degraded by white or black blotches seen in motion picture films prior to the 1950s have attracted attention. Conventional restoration methods are composed of two processes. The first process is motion estimation by block matching (BM). The second process matches the current frame to the previous and next frames based on the motion estimates and filters in the temporal direction. Blotch detection accompanies the filtering in the second process. The BM used in the first process is average BM or sorted BM. Blotch detection in the second process employs Spike Detection Index-a (SDIa). In conventional BM, however, there are limits to the accuracy of motion estimation when blotches are present in the blocks used to estimate the motion in images corrupted by blotches. Furthermore, the accuracy of the filtering in the second process is affected by the accuracy of the motion estimation. And although blotch detection based on SDIa is a simple method, the accuracy of its blotch detection is a problem. To solve these problems in this paper, we create a blotch position image by using a median filter with an iteration constraint as the first stage, and obtain a virtual original image with the blotches removed by filtering based on the blotch position image. The accuracy of the motion estimation is improved by using the virtual original image, and the accuracy of the filtering is improved by using the blotch position image in the filtering. Based on these innovations, we propose a new method to obtain a restored image sequence having enhanced image quality. The effectiveness of the proposed method is clearly demonstrated through a variety of examples. © 2002 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 85(10): 8–18, 2002; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.1121 [ABSTRACT FROM AUTHOR]

This paper introduces a peak enhancement method using an off-axis spectrum which is used in peak picking. It also considers its advantages and problems for the application of speech recognition. An off-axis model modified to be suitable for speech recognition is proposed. The modified off-axis model is realized by combining a peak emphasis model and a peak repressed model. The modified off-axis model can control the peak bandwidths and the slope of its spectrum by operating two parameters. The values of the parameters are determined adaptively to each sample by a stability test. In addition, some results of experimental speech recognition under noisy conditions are described and the efficiency of the method is discussed. [ABSTRACT FROM AUTHOR]

This paper discusses the speech recognition based on the time course of the local peak of the spectrum such as the formant, which has been considered important in the phoneme perception. A measure for the dynamical behavior of the spectrum is proposed based on the functional model of the FM-neuron which is shown to exist in auditory physiology. First, the FM-neuron is modeled as a time-frequency filter for the spectral time-series which responds only to the shift of the local peak frequency of the spectrum with the discrimination function for the shift direction. Then the measure to represent the difference of the output from the FM-neuron model is derived based on the cepstral expansion of the spectrum. The measure is called the spectral movement similarity. It is shown that the spectral movement similarity. It is shown that the spectral movement similarity on the auditory non-linear frequency axis can be realized equivalently by the frequency weighting. A spoken word recognition experiment is conducted employing the dynamic time warping (DTW) using the spectral movement similarity. It is shown also that the recognition error is reduced greatly by combining the proposed measure and the traditional spectral distance compared to the case where only the traditional spectral distance is used. This improvement is more remarkable when the spectral distance is used as the spectral distance, with the recognition error being reduced to one-fourth. [ABSTRACT FROM AUTHOR]

This paper proposes recognition methods that use hierarchical spectral dynamic features extracted over multiple time lengths. The effectiveness of these features in phoneme recognition and isolated word recognition is shown. The dynamic features are extracted for time sequences of both cepstral coefficients and logarithmic energy, and are combined with instantaneous cepstral coefficients. Input speech Is quantized by word- or phoneme-specific codebooks created by clustering a set of vectors consisting of these features for each word or phoneme. Speaker-independent isolated word recognition experiments are performed using a vocabulary of 100 Japanese words. When VQ distortion is used for word identification, the high recognition accuracy of 96 percent is achieved, and when VQ distortion is used for preprocessing, the number of word candidates for each input utterance is reduced to 1 percent of the vocabulary without increasing the error rate. Phoneme recognition experiments are performed far the /b/, /d/ and /g/ consonants in a large vocabulary of isolated words uttered by one male speaker. Using the proposed recognition method, the high recognition accuracy of 99 percent, which is similar to the accuracies obtained by conventional HMM or neural network approaches, is obtained. This paper also compares multiple codebook and single code-book methods. [ABSTRACT FROM AUTHOR]

This paper proposes a speech recognition system based on the pattern prediction using neural network. In the proposed system, an independent nonlinear predictor composed of a series of multilayer perceptrons (MLP) is prepared for each class which is the object of recognition. The temporal structure of the speech pattern, especially the temporal correlation structure between feature vector sequence, is represented by the nonlinear mapping between the input and the output, and is utilized as the important feature in the recognition. On the other hand, the variation of the temporal structure of the speech pattern, due to the difference of speakers and the fluctuation of the utterance, is normalized by the dynamic programming. As the training algorithm to determine the MLP parameters composing each predictor, an iterative algorithm combining the dynamic programming and the error backpropagation is proposed, together with the proof for the convergence. A speaker independent isolated digit recognition experiment is executed to examine the basic operation of the proposed system. The parameters are estimated in a satisfactory way even from a small number of training data, and it is indicated that a high recognition performance is realized. [ABSTRACT FROM AUTHOR]

This paper evaluates first the smoothed group delay spectrum (SGDS) distance measure through the isolated work speech recognition experiment by specified speakers. The experiment was performed for the following three cases, considering the speech recognition in the actual environment: 1) the case where the channels have difference characteristics; 2) the case where a white noise is added to the input speech; and 3) the case where the telephone speech is used as the input. In all three cases, the recognition rate is improved drastically compared to the traditional LPC cepstrum distance measure. An improvement of the recognition rate by 16 percent was realized under the noise of seg- mental SN ratio 20 dB. Then the distance measure is evaluated for the case where the FFT cepstrum is converted into the group delay spectrum. The proposed method gives a better recognition rate compared to the conventional FFT cepstrum distance measure, but the result is worse than the SGDA measure by approximately 3 percent since the higher-order FFT cepstrum coefficient has a larger variance on the time axis. Finally, the SGDS distance measure is evaluated by the isolated word speech recognition system with the monosyllable as the registered speech. The vowel recognition rate is improved, which improved the recognition rates for the syllable and the word by 2 percent or more on a relative scale. [ABSTRACT FROM AUTHOR]

This paper proposes a method for extraction of the spectral envelope by a spectral evaluation criterion with nonuniform weighting on a log magnitude scale. In the traditional cepstral method, the spectral envelope is obtained by linear smoothing on the log scale. Consequently, the obtained log magnitude spectrum is affected by the low-level portion of the spectral fine structure, one of the factors that has degraded the quality of the synthesized speech and the speech recognition rate. To avoid such a situation, the proposed method weights the peaks of the fine spectral structure more heavily than the valleys in the error criterion. Consequently, the resultant spectral envelope fits the peaks, i.e., the sampling points of the true spectral envelope f or a voiced sound. This method does not require pitch extraction. In addition, the spectral envelope can be obtained stably both for voiced and unvoiced sounds. Thus, it is suitable for the automatic analysis of speech. The effectiveness of the method is demonstrated by some examples, and its relation to other methods, especially to the improved cepstral method, is discussed. [ABSTRACT FROM AUTHOR]

This paper describes a new HMM synthesis method in which HMM adapts to additive noise and multiplicative distortion. The conventional HMM synthesis method can only be applied to additive noise. In the method described here, the likelihood of the synthesized HMM to adapt to speech is maximized, so that multiplicative distortion is eliminated when the method is applied. Within the framework of this method, adaptation to variations in the SN ratio, considered a problem in conventional HMM synthesis, can be formulated as part of the adaptation to multiplicative distortion. As a result of evaluating speech recognition rates using our method, we have confirmed that the method is effective for improving the recognition rate of speech that contains additive noise and multiplicative distortion. © 1998 Scripta Technica, Electron Comm Jpn Pt 3, 81(8): 1–9, 1998 [ABSTRACT FROM AUTHOR]

In this paper, a phonemic segmentation system is constructed which will be useful in practice as a prepro- cessing for English continuous speech recognition. The usefulness of the system is demonstrated. English continuous speech contains many weakly voiced vowels with ambiguous utterances called schwa, making stable detec- tion of the phonemic boundary difficult. In fact, there has been almost no detailed proposal for a segmentation system that can be used in practice as a preprocessing to the recognition. In the proposed method, the mel cepstrum is extracted from the speech signal, using the unbiased esti- mation of the log spectrum. This process is known to be stable, i.e., less affected by the fine spectral structure. Then the dynamic segmentation parameters are determined from the mel cepstrum such that the boundary of the schwa phonemes can be detected using the pseudo- differentiation filter. By combining the dynamic parameters with the static parameters, the phonemewise segmentation is executed in a hierarchical form. The proposed system can segment English continuous speech containing diversified phonemic environments into phonemes on the time axis. The system operates based only on the acoustic knowledge of phonemes, which are common to speakers with different qualities and utterances, without using the speaker-dependent complex boundary detection rules. Also, the proposed system is evaluated by experiment using English continuous phoneme-balanced speech uttered by one each native English-speaking male and female for 350 s. For the total number of phonemes 3024, the detect ratio for the phoneme boundary is 97.1 percent, the boundary delete ratio is 2.9 percent, and the boundary add ratio is 24.2 percent. [ABSTRACT FROM AUTHOR]

This paper proposes a speaker-independent spoken-word recognition system of the following structure. The speech segments at the phoneme level are obtained by the automatic segmentation of the speech on the time axis. The segments are labeled in parallel using multiple different phoneme reference pattern sets. The symbol sequence matching between the obtained multiple phoneme lattices and the dictionary items is attempted for each phoneme lattice, which results in the recognition of the word. The effectiveness of the method is verified by an experiment. To cope with the speaker with greatly different utterance modes and voice characteristics, a separate phoneme reference pattern set is prepared. By executing the phoneme labeling in parallel using such a pattern set, one of the phoneme reference pattern sets can always be selected to handle adequately any given speech of the speaker. In other words, the phoneme sequence is recognized at the stage of the word recognition. The processing to specify the reference pattern set matched to the input speaker from the multiple phoneme reference pattern sets is no longer required at the stage of the phoneme labeling. Based on the speech data uttered by three males and three females, six phoneme reference pattern sets are constructed for each speaker. Using those pattern sets, a spoken word recognition system is constructed by the parallel phoneme labeling method. The speaker-open speech recognition experiment is executed for the phoneme- balanced 212 words uttered by six males and six females. The result is that the average recognition rate for all speakers is 88.2 percent and the lowest recognition rate for the individual speakers is 83.0 percent. Compared to the system using the speaker-independent reference pattern set obtained by mixing the six phoneme reference pattern sets for all the male and female speakers, the proposed system has the average recognition rate higher by 1.9 percent and the lowest recognition rate for the individual speakers higher by 5.2 percent. In other words, a satisfactory result is obtained as the speaker independent spoke word recognition system. [ABSTRACT FROM AUTHOR]

This paper proposes a new vector quantization method which can reduce search complexity and code book memory size by reducing the number of code vectors without increasing quantization distortion. This method uses hypercolumnar clusters; and an input vector is quantized to a cluster, the center axis of which is nearest to the input vector. Thus, one vector and one scaler must be coded and transmitted. The proposed method was applied to four geometrically different distributions and LPC cepstra of speech signals. As a result, the number of code vectors was decreased comapared with that in an ordinary vector quantizer, in all of the forementioned distributions. That the reduction of memory size and search complexity were evaluated. [ABSTRACT FROM AUTHOR]

In speech recognition, speaker adaptation refers to the range of techniques whereby a speech recognition system is adapted to the acoustic features of a specific user using a small sample of utterances from that user. In recent years the practical development of speaker-independent speech recognition systems using continuous density hidden Markov models has seen significant progress; however, the recognition performance of these systems has not yet reached that of speaker-dependent speech recognition systems in which a user's speech is registered beforehand. Much hope has therefore been placed on the establishment of speaker adaptation techniques that can bring performance of a speaker-independent system up to that of a speaker-dependent one using the smallest amounts of data. In this paper we present a survey of previous research into speaker adaptation techniques focusing particularly on three important approaches in this area: maximum a posteriori (MAP) parameter estimation, maximum likelihood linear regression (MLLR), and eigenvoices. We also discuss approaches that combine these techniques in a lateral fashion. © 2005 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 88(12): 25–42, 2005; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ ecjc.20207 [ABSTRACT FROM AUTHOR]

According to the boundary surface control principle, it is possible to reflect the flow of acoustic energy at a boundary if a boundary surface with impedance 0 can be formed by controlling the sound pressure and the particle velocity at an arbitrary position in space. If this principle is used in such a way that a boundary surface is formed only in the direction not intended for radiation of the sound, then a directional speaker system can be realized. In the conventional design method of the directional speaker array system, in order to achieve sharp directivity in the low frequencies, it is necessary to enlarge the array length. On the other hand, in our proposed system, it is easy to achieve sharp directivity in the low frequencies, since the boundary surface control principle can achieve high performance in the low frequencies. Moreover, by changing the control sound source spacing for each bandwidth, rather flat characteristics can be obtained over the entire control range. In this paper, the basic characteristics constituting guidelines for system design and the effect of control are confirmed by numerical calculations and experiments. © 2004 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 88(2): 1–9, 2005; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.20122 [ABSTRACT FROM AUTHOR]

In speech recognition systems where the speaker and utterance environment cannot be designated, the drop in recognition precision due to the incompatibility of the input speech and acoustic model's training data is a problem. Although this problem is normally solved by speaker adaptation, sufficient precision cannot be achieved for speaker adaptation unless good-quality adaptation data can be obtained. In this paper, the authors propose a method of efficiently clustering large-scale data using the likelihoods of a cluster model that was created from small-scale data as the criteria to obtain a high-precision adapted acoustic model. They also propose a method of using the cluster model to automatically determine the adapted acoustic model during recognition from only the beginning of the sentences of the input speech. The results of applying the proposed technique to news speech recognition experiments show that the adapted acoustic model selection precision can be ensured by using only 0.5 second of data of the beginnings of sentences of the input speech and that the proposed technique achieves a reduction rate for invalid recognitions of 20% and a reduction in the time required for recognition of 23% compared with when the adapted acoustic model for each cluster is not used. © 2004 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 88(2): 41–51, 2005; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.20157 [ABSTRACT FROM AUTHOR]

This paper discusses the induced effect on the body sway when an iteratively rotating visual wide-field stimulus and an auditory stimulus with a moving sound image are presented simultaneously with synchronization. When the sound image is iteratively rotated in phase relative to the iteratively rotating visual stimulus, the induced effect on the body sway increases when the visual angle of the visual stimulus is increased, compared to the case in which the auditory stimulus is presented in the opposite phase. When the visual angle of the visual stimulus is reduced, on the other hand, there is no difference in the induced effects between the in-phase and opposite-phase conditions. These results suggest that when the visual and auditory stimuli are simultaneously presented, the induced effects of the visual and auditory stimuli on body sway are not simply added, but that there is an interaction between the two effects. It is suggested that the induced effect due to the interaction between the visual and auditory stimuli depends on the visual angle of the visual stimulus, and is enhanced by the wider field visual stimulus. © 2002 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 85(10): 31–39, 2002; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.1123 [ABSTRACT FROM AUTHOR]

This paper presents a study for improving the sensory substitution system using the fingertip tactile sensation as an auxiliary means of lipreading and talking with hands for the severely hearing impaired, for whom the hearing aid or the cochlear implant was not effective. First, the problems of the previously developed 16-channel tactile vocoder with a vibrator array are described. It is pointed out that the backward masking of the tactile sensation is one of the factors that made consonants difficult to recognize. Then, as a solution to the problem, a new stimulus presentation method is proposed where the speech spectrum is swept on the finger surface. The optimal sweeping velocity in determined from the viewpoint of the two-point threshold so that the backward masking is minimized. Based on those investigations, an 8 × 4 vibrator matrix is constructed which can sweep the speech spectrum. A recognition experiment is attempted for monosyllabic speech. It is seen as a result that the identification of a consonant that is easily affected by the backward masking, especially plosives are improved. Thus, it is suggested that the sweeping of the spectral pattern is effective as an auxiliary means of speech recognition, even for the profoundly hearing impaired. [ABSTRACT FROM AUTHOR]

This paper proposes a speech recognition algorithm for large vocabulary continuous speech. The proposed algorithm is based on the hidden Markov model (HMM)- LR algorithm using a generalized predictive LR parser and phoneme HMMs. The following three techniques are applied to improve recognition performance and reduce processing time. (1) The forward and the backward likelihood are used to accurately determine the likelihood in the beam search. (2) To reduce the trellis computation in HMM speech recognition and for efficient search, only the speech frames in which the predicted phoneme seems to exist are used by the window for phoneme matching. (3) For efficient search, adjusting identical phoneme sequences are merged by checking the stack and the state of the LR parser. The algorithm was applied to a telephone directory assistance task involving more than 70,000 subscribers. A recognition experiment for continuous word utterance was done. The sentence recognition rate was 85 percent speech recognition. The sentence understanding rate was 59 percent for speaker-dependent speech recognition with spontaneous utterances. [ABSTRACT FROM AUTHOR]

This paper describes the speaker-independent spontaneous speech dialog system TOSBURG II, where the user can input the speech interrupting the speech response from the system. In the speech dialog system TOSBURG, which was previously constructed experimentally, spontaneous utterance dialog was realized by a continuous speech understanding based on the key word and the multimode response. TOSBURG II aims at dialog that is more flexible and interactive. The speech response component superposed on the input speech signal is cancelled by the active control technique, making it possible for the user to interrupt and input during the system speech response. To improve the precision of speech recognition, noise immunity learning, which is useful in detecting the keyword in the noise, is extended so that it can also be applied to the speech input after the speech response cancellation, where the residual component still remains. To evaluate the effectiveness of speech response cancellation and noise immunity learning, an experiment is executed by superposing the speech data (obtained by the utterance of 350 sentences containing 49 key words by five adult males) on the system speech response. The key word detection rate, false alarm rate, and sentence recognition rate are improved from 58.8 percent, 24.1 FA/H/ W (false alarms/hour/word), 27.8 percent, to 89.6 percent, 15.4FA/HAV, 58.4 percent, respectively. A dialog experiment is also conducted using the constructed realtime system, and the usefulness of the proposed system is verified. [ABSTRACT FROM AUTHOR]

This paper considers the codebook for low bit-rate speech coding by code-excited linear prediction (CELP) and the mapping table for the channel code. It also proposes an optimization procedure for sequential training based on the structured excitation codebook. In the discussions of structurization of the codebook, the conjugate structure to represent the output by the sum of two vectors, the pitch synchronized processing, rotating use of the codebook, the time-domain sloped gain, and the integrated mapping to the channel code are considered. The distortion measure used in the training is the same closed-loop measure as is used in the coding. In other words, the distortion between the input speech and the synthesized speech at the perceptually weighted filter output is minimized. By training the codebook, the SNR of the coded speech is improved by approximately I dB. The improvement in speech quality is also verified. It is also verified that the robustness against bit error can be improved by the training of the mapping table to the channel code, as well as by training based on the distortion measure considering bit error. [ABSTRACT FROM AUTHOR]

This paper proposes a techoique for supervised speaker adaptation for speech recognition based on a continuous density hidden Markov model (HMM). When the number of utterances for the adaptation words is decreased in order to reduce the burden on the user, the ratio of demisyllables (recognition units) not appearing in the uttered words is increased. In the proposed method, the HMM parameters for the demisyllables contained in the adaptation data are corrected and then the HMM parameters for the demisyllables not contained in the adaptation data are corrected by interpolation in the parameter space (spectral interpolation). To avoid a biased estimation of the parameters that depend on the adaptation data set, a correction is made based on largescale uttered data by a large number of speakers. A word recognition experiment, executed for 100 similar words, simulates a 5000-word large-vocabulary speech recognition; and the proposed method is then evaluated. When the recognition rate by the speaker-independent HMM reaches 81.2 percent, the recognition rate is improved up to 85.2 percent by using 50 words for adaptation. Thus, the effectiveness of the proposed method is verified. [ABSTRACT FROM AUTHOR]

This paper proposes a method to detect the prosodic phrase for speaker-independent continuous speech. The method uses continuous adjustment of a pitch pattern with the learning accent pattern (pitch template). The accent phrase with one accent cue and its corresponding pitch pattern are considered correlated. Distinctive pitch templates are obtained by classifying the pitch patterns corresponding to the accent phrase with the help of clustering. Since pitch patterns of continuous speech are connected to these pitch templates, prosodic phrase segmentation can easily be done by One-Stage DP matching of the pitch pattern with pitch templates. The method can also be applied to unspecified speakers whose average pitch frequency differs from each other by sliding the pitch template along the vertical axis during the matching. The ATR continuous speech database (10 speakers, 503 sentences) is used for the experiments. For eight templates, about 65 percent of the automatic prosodic phrase segmentations were correct and the experiment was successful in detecting about 83 percent of the prosodic phrase boundaries. [ABSTRACT FROM AUTHOR]

Pitch frequency is one of the most important voice characteristics, and its accurate extraction is important not only in speech analysis and synthesis, but also in speech coding, speech recognition, speaker recognition, and the like. Existing methods of improving extraction accuracy include waveform processing, correlative processing, and spectral processing. This paper describes the use of a neural network to extract pitch from voice features delivered from the bandpass filter pairs (BPFPs) proposed by Fonda et al. Three types of multi-layered neural networks able to learn time-continuity and high accuracy discrimination functions and have a recurrent structure are tested. The cross-coupling multi-layered neural network with feedback architecture gives the best improvement over conventional neural networks, and exhibits superior ability for learning time continuity of pitch and U/V information. © 1997 Scripta Technica, Inc. Electron Comm Jpn Pt 3, 80(9): 48–58, 1997 [ABSTRACT FROM AUTHOR]

Speech recognition has been studied extensively in recent years with regard to human interface and the goal has been realization of speech recognition systems. However, the problem is that these systems have numerous limitations, despite being accompanied by implementation of speech-based interactive systems. There have been a few attempts to use associative information in Japanese speech recognition and understanding. As a result, this paper introduces a new concept of association between semantic attributes and associations between words under the assumption of so-called “indirect associations” in human communication. Hence, nearly successful processing of associations close to common sense relations is achieved. © 1997 Scripta Technica, Inc. Electron Comm Jpn Pt 3, 80(2): 63–77, 1997 [ABSTRACT FROM AUTHOR]

The results of speaker recognition methods using vector quantization (VQ) distortion and discrete or continuous ergodic hidden Markov models (HMMs) are compared. The effectiveness of these methods is examined from the viewpoint of robustness against utterance variation such as differences in content, temporal variation, and changes in utterance speed. It is shown that the continuous HMM performs much better than the discrete HMM and its performance is close to that of the VQ distortion method. When the amount of training data is limited, however, the VQ distortion method achieves a better recognition rate than the continuous HMM. The transition information between the states is shown to contribute little to identifying the individual characteristics of a voice. An increase in the number of states or in the number of mixture components in a state both have an equal effect, and recognition performance is almost completely determined by the product of these two numbers. [ABSTRACT FROM AUTHOR]

We have investigated methods for reducing the number of word hypotheses registered in the word graph and the amount of memory used by the factoring tables for the tree-structured dictionary with the objective of reducing the memory requirements of a continuous speech recognition system. By assigning word hypotheses in the word graph attributes relating to the number of continuation hypotheses in which they are included, we are able to efficiently determine unusable word hypotheses during pruning and can perform garbage collection. This procedure allows us to reduce the amount of memory needed for generating word hypotheses from 127 MB to 6.9 MB. In addition, by approximating the bigram values held in the factoring tables with POS bigrams, we were able to reduce the memory consumption of the factoring tables from 56 MB to 19 MB with almost no impairment of recognition performance. As a result of these reductions in memory requirements, the memory consumption of the decoder has been reduced from 246 MB to 113 MB. © 2005 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 89(2): 54–64, 2006; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.20221 [ABSTRACT FROM AUTHOR]

Now that dictation systems have become a reality, research on large-vocabulary voice recognition is shifting from “read speech” to “spontaneous speech.” The features of this kind of spontaneous speech have been investigated from various perspectives using a corpus of dialog speech. However, in particular for the Japanese language, creating a voice recognition system which has massive amounts of data and is based on statistical methods is not necessarily sufficient. There have been few reports on the performance of a large-vocabulary voice recognition system whose purpose is tagging spontaneous speech. The authors have prepared a spontaneous speech corpus using as materials lecture speeches from the University of the Air for the purpose of improving the recognition precision for spontaneous speech. They report here on the recognition performance of their large-vocabulary spontaneous speech recognition system created using this corpus. The results of experiments showed that the word error rate for lecture speeches, which was 51.5% under conventional systems that use reading, was reduced to 16.4%. © 2003 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 86(8): 52–60, 2003; Published online in Wiley InterScience (www. interscience.wiley.com). DOI 10.1002/ecjc.10105 [ABSTRACT FROM AUTHOR]

A method for handling noise in a car information machine incorporating speech recognition middleware and having a SuperH (SH) Micon as its platform is reported. Using both a noise acoustic model and a spectral subtraction scheme, this method is evaluated by computer simulation against an evaluation database constructed by considering the usage environment and the microphone position. In evaluations using the noise model alone, a speech recognition rate of 11.0% in a test track driving environment was obtained when an HMM tutored by speech without noise was used, while a recognition rate of 82.7% was obtained when an HMM constructed by superimposing noise over the tutoring speech was used. The speech recognition rate was improved by about an additional 5%, to 87.6%, by using this noise-containing acoustic model and the spectral subtraction scheme. In addition, when speech recognition middleware incorporated with the antinoise method was loaded on an SH board, a recognition performance comparable with the computer simulation results was obtained. © 2002 Wiley Periodicals, Inc. Electron Comm Jpn Pt 3, 85(11): 65–73, 2002; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjc.10055 [ABSTRACT FROM AUTHOR]

The smoothed group delay spectrum distance (SGDS) measure is evaluated in speaker-independent recognition experiments. First, the appropriate level of smoothing of the group delay spectrum (GDS) is investigates by adding noise, etc., to the input speech. Then a comparison with the speaker-dependent case is made. An experiment is reported in which, for low amplitude parts of speech (e.g., unvoiced speech), the standard (LPC) distance measure is used in the interframe distance calculation instead of the SGDS distance measure. This method prevents a loss of recognition accuracy due to too strong an emphasis on certain spectral elements and a consistently high recognition accuracy can be achieved. Finally, evaluate the SGDS distance measure is evaluated where the GDS is represented in the spectral domain as a discrete Fourier transform (DFT) of the LPC coefficients. In comparison to the SGDS which was calculated by weighting the LPC cepstrum coefficients, computation time and memory space can be reduced without loss of recognition accuracy. Furthermore, a low bit quantization of the GDS is reported and a high recognition rate is achieved with only 32 bits per frame. [ABSTRACT FROM AUTHOR]