Search

The proposed system used stereo type depth camera by examining the human action recognition and also sleep monitoring in the elderly care center. Different regions of interest (ROI) are extracted using the U-Disparity and V-Disparity maps. The main information used for recognition is 3D human centroid height relative to the floor and percentage of movement from frame differencing for sleep monitoring. The results from the experiments of the proposed method show that this system can detect the person location, sitting or lying and also sleep behaviors effectively.

We investigated a method to detect the position and direction of travel of a pedestrian walking indoors simultaneously. Multiple time instances of Received Signal Strength Indicator from multiple Bluetooth Low Energy beacons, detected on smartphones carried by pedestrians moving in any of eight directions, are used to train a Deep Neural Network to estimate the location and direction of motion of this smartphone. We compare the estimation accuracy when the RSSIs were interpolated, and when they were not. The position estimation accuracy without interpolation decreased by 40.5% and the moving direction estimation accuracy decreased by 26.9%, suggesting some form of interpolation is required.

Reverberation distorts speech leading to degradation of speech intelligibility and speech quality. We have been investigating methods to predict speech intelligibility without clean speech which are more difficult than prediction with actual clean speech. In this paper, we attempt to estimate speech intelligibility by simulating the full-reference (F-R) method, substituting the predicted clean speech with clean speech. The correlation between the predicted and measured intelligibility was 0.99945 and the RMSE was 0.00353 with the proposed method.

In this research, we examined whether it is possible to present sound in areas (audio spots) with various shapes and sizes using Flexible Parametric Array Loudspeakers (FPALs). In addition to multiple Parametric Array loudspeakers, an audio spot was created using FPALs that can change the beam shape of a conventional Parametric Array Speaker (PAL) by changing its curvature. Changes in the directivity pattern of the audio spots created with FPALs were evaluated by measuring sound pressure in the area in front. As a result, it was confirmed that the shape of the audio spot can be controlled as intended.

We investigated a method to simultaneously detect the location and motion direction of a pedestrian walking indoors. Multiple time instances of Received Signal Strength Indicator (RSSI) readings from multiple Bluetooth Low Energy (BLE) beacons on a smartphone held by a pedestrian, traveling in one of 9 directions, were fed to a trained Deep Neural Network (DNN), and the location of this smartphone, as well as the direction of its motion, was simultaneously estimated. Previous experiments in a $4m\times 7m$ area showed estimated location accuracy of 0.91m, and average estimation accuracy of 83.5% in 5 directions. In this paper, we significantly increased the quality of the RSSI training data, as well as the number of motion direction to nine from five. The estimated location accuracy of 0.439m, and average estimated direction accuracy of 81.2% in 9 directions was shown.

We proposed and evaluated a speech intelligibility estimation method for binaural signals. The assumption here was that both the speech and competing noise are directional sources. In this case, when the speech and noise are located away from each other, the intelligibility generally improves since the auditory system can segregate these two streams. However, since intelligibility tests as well as its estimation is conducted based on monaurally-recorded signals, this potential increase in the intelligibility due to the segregation of sources is not accounted for, and the intelligibility is often under-estimated. Accordingly, in order to estimate the intelligibility taking into account this binaural advantage, we trained a mapping function between the subjective intelligibility and objective measures that account for the binaural advantage stated above. We attempted SNR calculation on (1) a simple binaural to monaural mix-down, which models the conventional estimation, (2) simple pooling of both binaural channels (pooled channel), (3) channel signal selection with the better SNR from left and right channels (better-ear), and (4) sub-band wise better-ear selection (band-wise better-ear). For the mapping function training, we tried neural networks (NN), support vector regression (SVR), and random forests (RF), and compared these to simple logistic regression (LR). We also investigated the sub-band configuration that gives the best estimation accuracy by balancing the frequency resolution and the amount of training data. It was found that the combination of the better-ear model and RF gave the best results, with root mean square error (RMSE) of about 0.11 and correlation of 0.92 in an open set test.

Smart technologies are necessary for ambient assisted living (AAL) to help family members, caregivers, and health-care professionals in providing care for elderly people independently. Among these technologies, the current work is proposed as a computer vision-based solution that can monitor the elderly by recognizing actions using a stereo depth camera. In this work, we introduce a system that fuses together feature extraction methods from previous works in a novel combination of action recognition. Using depth frame sequences provided by the depth camera, the system localizes people by extracting different regions of interest (ROI) from UV-disparity maps. As for feature vectors, the spatial-temporal features of two action representation maps (depth motion appearance (DMA) and depth motion history (DMH) with a histogram of oriented gradients (HOG) descriptor) are used in combination with the distance-based features, and fused together with the automatic rounding method for action recognition of continuous long frame sequences. The experimental results are tested using random frame sequences from a dataset that was collected at an elder care center, demonstrating that the proposed system can detect various actions in real-time with reasonable recognition rates, regardless of the length of the image sequences.

Objective measures of intelligibility are preferable to subjective ones in the evaluation of speech systems used in real environments. In this study, subjective evaluations of eight types of indoor noise environments were used to compare four intelligibility indices to objectively evaluate Japanese speech intelligibility. These indices were as follows: short-time objective intelligibility (STOI), which has been widely used in recent years; speech intelligibility prediction based on mutual information (SIMI), which is derived from STOI; extended STOI (ESTOI), which is an improved version of STOI; and frequency weighted segmental signal to noise ratio (fwSNRseg), which incorporates both time and frequency components. These indices were subjectively evaluated in the eight noisy environments included in the corpus and environments for noisy speech recognition 4 (CENSREC-4) dataset using the familiarity-controlled word lists 2007 (FW07) as the speech data for the intelligibility evaluations. The results of the subjective evaluation of the four indices were then used to train predictive intelligibility estimation models. We evaluated the model performance using cross validation, which involved repeated training of seven of the eight environments and predicting the speech intelligibility under the remaining one environment. In the simulation results, the prediction accuracy of the SIMI index was significantly higher than that of the other indices, with a root mean squared error of 0.160 and a correlation coefficient of 0.934.

We previously studied DNN as a method for reducing wind noise in bicycles AAR navigation systems. As a result, DNN was able to remove almost all of the wind noise in most cases. However, there were cases where wind noise was not reduced sufficiently when wind noise and environmental noise were simultaneously generated. Another problem was that the timbre of the environmental sound changed and the original quality could not be maintained. In this paper, we evaluated wind noise reduction and environmental sound that passes through using newly trained DNN. As a result, while the wind noise removal performance remained very high, the timbre alteration of the environmental noise was kept to a minimum.

In this study, we clarified how much the amount of snow accumulation, the amount of snowfall and the density, the major factors of snow, contribute to speech intelligibility using multivariate analysis. In order to measure the transfer characteristics of the snow, an impulse response measurement experiment was performed 13 times under different conditions outdoors in the snow. The measured transfer characteristics were convoluted with the test speech, and the Japanese intelligibility test was performed using these samples. Multiple regression analysis was performed using speech intelligibility as the objective variable and snow factor as explanatory variables, and the contribution rate to intelligibility was analyzed.

We have been investigating the possibility of controlling noise using active noise control (ANC) technology. We proposed a novel ANC structure which uses Audio spotlights (AS) with subdivided AM sideband as the secondary source. The secondary sound source, which cancels noise, is only generated within the spotlight. We compared the efficiency of the proposed ANC with ANC that uses a conventional parametric array speaker (PAL) that delivers secondary source in a beam, and found that the proposed ANC can cancel noise only in a limited area, while additional noise is not introduced in the surrounding area.

We proposed a manipulation detection method for interrogation speech. We used a robust fingerprinting method optimized for speech since our intended target is interrogation speech recorded during a police investigation. The fingerprint uses line spectral pairs (LSP) to measure the spectral envelope of the speech, and is coarsely quantized so that the fingerprint will not be altered by small degradation in the signal, but will be altered enough by malicious modifications to the speech content. This fingerprint is embedded in the speech signal using conventional spread-spectrum watermarks. To detect manipulation, the watermarked fingerprint is detected, and compared to the fingerprint extracted from the speech itself. If the fingerprints match within the predetermined tolerance, it can be authenticated to be unaltered. Otherwise, manipulation should be suspected.

We have been studying objective evaluation methods of speech intelligibility. In order to use the full reference method, which can estimate the speech quality with high accuracy, we propose a method to estimate the clean speech, which becomes the reference signal, from the noisy speech by noise removal. We have constructed a Deep Neural Network to eliminate noise and evaluated its noise removal performance.

Reverberation distorts speech leading to degradation of speech intelligibility and speech quality. We have been investigating on methods to predict the speech intelligibility of reverberant speech. This normally requires clean speech, which is often not available. In this paper we attempt to use Deep neural networks (DNNs) to reconstruct clean speech from distorted speech. This approach proved to be effective for dereverberation in various reverberant environments.

We have been studying a navigation system for the visually impaired using 3D audio. In previous research, simulation experiments were carried out using prototypes and evaluated. However, since this simulation does not let the users actually move around, it was not able to verify possible problems in actual use. In this research, we aim to evaluate the 3D audio navigation system by implementing it on a smartphone and actually having the users move along the specified paths.

Various real time noise reduction methods that can actually be used in bicycles are systematically evaluated and compared, and the optimum wind noise countermeasure is examined. Accordingly, the effect of wind noise reduction by some wind noise reduction method was evaluated in terms of the intelligibility of the augmented navigation speech. In addition, we evaluated the influence of these measures on the environmental sound direction perception accuracy, since this is critical for detection of potential hazards. As a result, the Wiener iteration filter was found to be the most effective for wind noise reduction, but the combined use of a high pass filter and a windshield was also a well-balanced alternative method.

We have been studying a navigation system for the visually impaired using 3D audio. We proposed to give cues on the direction and distance to the target location using a simple 3D guidance audio. We implemented a prototype 3D audio navigation system, and conducted experiments using this prototype. We investigated the influence that sound pressure level differences of navigation sound gave for a route guidance.

We proposed a manipulation detection method for interrogation speech. We used a robust fingerprinting method optimized for speech since our intended target is interrogation speech recorded during a police investigation. The fingerprint uses line spectral pairs (LSP) to measure the spectral envelope of the speech, and is coarsely quantized so that the fingerprint will not be altered by small degradation in the signal, but will be altered enough by malicious modifications to the speech content. This fingerprint is embedded in the speech signal using conventional spread-spectrum watermarks. To detect manipulation, the watermarked fingerprint is detected, and compared to the fingerprint extracted from the speech itself. If the fingerprints match within the predetermined tolerance, it can be authenticated to be unaltered. Otherwise, manipulation should be suspected. We conducted initial experiments to verify the feasibility of the proposed method, and confirmed that at the utterance level, we can identify all substitution manipulated speech utterances successfully.

We evaluated the effect of synchronization on audio watermarks. Synchronization signal was generated by using an M-sequence for the phase, and shaping its magnitude relative to the host signal spectrum. Synchronization at the detector was achieved by calculating a cross-correlation function with the same M-sequence, and detecting the peak in the correlation. The synchronization signal was added to a spread-spectrum watermarked audio signal. This signal was then transmitted over analog channels, by simply connecting the D/A output to the A/D input using a cable, and also playing out the analog signal through a loudspeaker and recording this signal using a microphone. The addition of the synchronization signal enabled the detection of the watermark through analog channels at acceptable accuracy. The synchronization rate was also shown to strongly influence the watermark bit error rate (BER).

We proposed and evaluated a speech intelligibility estimation method for binaural signals. The assumption here was that both the speech and competing noise are directional sources. We trained a mapping function between the subjective intelligibility and some objective measures. We attempted SNR calculation on a simple binaural to monaural mix-down, better SNR selection from left and right channels (better-ear), and a sub-band wise better-ear selection (band-wise betterear). For the mapping function training, we tried neural networks (NN), support vector regression (SVR), and random forests (RF). A combination of better-ear and RF gave the best results, with root mean square error (RMSE) of about 4% and correlation of 0.99 in a closed set test.

We studied the feasibility of the use of iBeacons for position detection. We compared the precision achieved by two well-known algorithms, i.e. nearest neighbor and k-nearest beacon. The nearest neighbor algorithm showed lower average error distance, by 1.0 m, than the k-nearest beacon algorithm. We also found that both indoor position methods are not influenced by the difference in the measured rooms. This accuracy should be enough for its application to navigation system using augmented audio reality, to be implemented in the future.

We have been studying a navigation system for the visually impaired using 3D audio. We proposed to give cues on the direction and distance to the target location using a simple 3D guidance audio. We implemented a prototype 3D audio navigation system, and conducted experiments using this prototype. We found that this system is effective in guiding the subjects through a complex path.

We evaluated the masking performance of the prototype super-directive masking system that uses a parametric speaker to steer a directional masker beam towards the head of the subject to be masked (maskee). The parametric speaker is placed on an iRobot-like platform (turtlebot), and tracks the maskee using the Kinect sensor. We compared the masking performance (intelligibility decrease) of the prototype system with a conventional masking system using loudspeakers, and found them to be comparable. Moreover, the prototype maintains the range of the masking noise to a narrow beam, thereby not annoying surrounding listeners with increased noise levels.

We have been investigating on the possibility of controlling singing voice using active noise control technology. We proposed a novel structure which introduces a widened portion in the pipe structure. We compared the efficiency of ANC with conventional conduction path pipes, and found that the new structure enhances the ANC performance depending on the phonetic content.

We investigated on the possibility of an active cancellation system for unnecessary speech radiation control. Some examples of the intended application of this system are cellular speech cancellation and speech input for recognition-based dictation systems. Both of these applications do not require speech to be radiated into surrounding space, but only into the input microphone, and would benefit if global radiation is controlled. We first show that speech cancellation is possible with a secondary source placed in proximity to the mouth generating linear-predicted phase-inverted speech. However, the prediction must also cover the long delay associated with the acoustic to/from electric conversion, as well as A/D, D/A conversions, and all associated processing, which we found could go up to as long as 3ms. By using LPC predicted samples recursively to predict further samples, we found that prediction with SNR of about 6dB is possible, even with this long delay. The prediction coefficient update is suppressed during this recursion. Lowering the sampling frequency in order to lower the number of predicted samples at the cost of reduced bandwidth further enhances prediction accuracy. At a sampling frequency of 8kHz, speech emission control of about 7dB for female speech and 4dB for male speech was found to be possible. Finally, we experimentally evaluated the proposed active speech control method. Predicted samples of recorded speech was first prepared off line. We then actually played out both the original and the predicted samples simultaneously from two loud speakers. It was found that (1) speech cancellation of up to about 10dB is possible, but is highly speaker dependent, (2) secondary loud speaker should be oriented in the same direction as the primary source, i.e., the mouth. We plan to investigate further to improve prediction accuracy using prediction coefficient extrapolation. A prototype system implementation using DSPs is also planned.

Most existing objective intelligibility prediction methods predict monaural intelligibility using monaural signals. These methods do not consider that a human can easily distinguish sounds arriving from different directions by sound heard in both ears. Therefore, intelligibility prediction using binaural signals that take this into account is necessary. Accordingly, speech samples with various source locations were prepared using binaural simulation. Subjective measurement tests were carried out on these samples. The frequency-weighted segmental SNR (fwSNRseg) was obtained using two different models for comparison: 1) a monaural model, which simply combines signals in both channels into one, and 2) the simple better ear model which uses the channel with the better SNR. Binaural intelligibility is estimated by applying regression analysis on the fwSNRseg and the subjective measurement results. Intelligibility was predicted by applying the resultant regression function on fwSNRseg of the test sample. We compared the estimation precision of the two models with binaural signals. The estimation precision of the better ear model yielded higher correlation with subjective scores than that of the monaural model by approximately 0.2 in a closed test as well as two open sets.

In this paper, we compared a number of frame synchronization methods for audio watermarks. M-sequence was embedded into the phase of the host signal for synchronization. We compared the detection accuracy of the Cross-power Spectrum Phase Analysis (CSPA) and the Smoothed Coherence Transform (SCT) method for additive noise. The CSPA showed higher detection rates by about 30% when 5-point smoothing was applied to the correlation. We plan to test other deterioration including analog channels using the CSPA.

We investigated the feasibility of using a head-tracking parametric speaker system to steer a directional masker beam towards the head of the subject to be masked (maskee). The parametric speaker is placed on an iRobot-like platform, and tracks the maskee using the Kinect sensor. We investigated the tracking performance using a dummy head, and found it to be better than an actual human by about 10%. Mask samples will be recorded using the dummy head, and presented to listeners using headphones for masking evaluations. We will have results of the masking performance evaluation by the conference date.

We have been investigating on the possibility of controlling singing voice using active noise control technology. We proposed a novel structure for conduction pipes with absorption material lining into which the singer can sing into. We compared the efficiency of ANC with conventional conduction path pipes, and found that the new structure enhances the ANC performance depending on the phonetic content.

We proposed and evaluated a speech packet loss concealment method which predicts lost segments from speech included in packets either before, or both before and after the lost packet. The lost segments are predicted recursively by using linear prediction both in the forward direction from the packet preceding the loss, and in the backward direction from the packet succeeding the lost segment. Predicted samples in each direction are smoothed by averaging using linear weights to obtain the final interpolated signal. The adjacent segments are also smoothed extensively to significantly reduce the speech quality discontinuity between the interpolated signal and the received speech signal. Subjective quality comparisons between the proposed method and the the packet loss concealment algorithm described in the ITU standard G.711 Appendix I showed similar scores up to about 10% packet loss. However, the proposed method showed higher scores above this loss rate, with Mean Opinion Score rating exceeding 2.4, even at an extremely high packet loss rate of 30%. Packet loss concealment of speech degraded with G.729 coding, and babble noise mixed speech showed similar trends, with the proposed method showing higher qualities at high loss rates. We plan to further improve the performance by using adaptive LPC prediction order depending on the estimated pitch, and adaptive LPC bandwidth expansion depending on the consecutive number of repetitive prediction, among many other improvements. We also plan to investigate complexity reduction using gradient LPC coefficient updates, and processing delay reduction using adaptive forward/bidirectional prediction modes depending on the measured packet loss ratio.

In this paper, we propose an “ANC mute microphone” to attenuate noise caused by singing voice. To verify that an ordinary Filtered-X LMS Algorithm is effective in cancelling singing voice, we conducted simulations and experimented using the DSP kit DSK6713. The simulations show that the attenuation of a singing voice is modestly possible (up to 4 dB). According to the real-time experiment using the DSK, low frequency voice, such as Bass, was attenuated relatively well, by over 8 dB. On the basis of these results, we plan to further improve this device for realistic applications.

In this paper, we describe a prototype human head tracking parametric speaker system that can present speech in a highly directional beam. This system is composed of the Kobuki robot platform, the Kinect sensor, a USB-controlled camera pod and a parametric speaker. We found that it is possible to track and deliver speech to a human walking at an average speed of 5 km/h. Experimental results have shown that it is possible to track and continuously present speech at a constant level to a walking subject, with less than 5 dB level variation.

In this paper, we compared the Linear Discriminant Analysis (LDA) with Random Forest (RF) for musical instrument identification from clips with a mixture of instruments. As the first step, monotone samples from the Musical Instrument Samples (Univ. Iowa) and RWC Music Database were used to identify the individual instruments. For the Iowa monotones, an overall instrument recognition rate of 24.8% and 82.1% was obtained using LDA and RF, respectively. However, the rate degrades to 54.9% on the RWC monotones even with RF, most likely due to insufficient number of features to cover the increase in variability of this large database.

We propose a voice picking system that uses localized guide speech oriented towards the target object, and Head Tracking (HT), which maintains the localized speech position according to the user's head movement. We verified the usability improvements achieved by the proposed system. The task completion time was reduced by about 1.8 seconds (11.2%) compared to conventional diotic conditions. The post test questionnaire showed that the HT condition is preferred over conventional diotic systems by 1.92 on a 7-point scale. Thus, with the introduction of sound localization in voice picking systems, both the user satisfaction and task efficiency can be improved.

A common problem in speech processing is the conversion of the written form of a language to a set of phonetic symbols representing the pronunciation. In this paper, we focus on an aspect of this problem specific to the Japanese language. Written Japanese consists of a mixture of three types of symbols: Kanji, Hiragana, and Katakana. We describe an algorithm for converting conventional Japanese orthography to a Hiragana-like symbol set that closely approximates the most common pronunciation of the text. The algorithm is based on two hypotheses: (1) the correct reading of a Kanji character can be determined by examining a small number of adjacent characters and (2) the number of such combinations required in a dictionary is manageable. The algorithm described here converts the input test by selecting the most probable sequence of orthographic units (n-grams) that can be concatenated to form the input text. In closed-set testing, the n-gram algorithm was shown to provide better performance than several public domain algorithms, achieving a sentence error rate of 3% on a wide range of text material. Though the focus of this paper is written Japanese, the pattern matching algorithm described here has applications to similar problems in other languages.

In this paper, we investigated on an efficient masker for speech privacy protection. Previously, we proposed and evaluated a speaker-dependent masker which is created from speech of the speaker being masked. This masker was proven to lower the intelligibility of speech at a much lower level than conventional babble maskers. However, this masker needs to be prepared for each potential speaker, and may be expensive to create and maintain. Accordingly, we investigated on an alternative masker, where the speech is masked by a masker created from a different speaker of the same gender. We measured the effectiveness of this gender-dependent masker, and found that male gender-dependent masker is feasible, with efficiency almost equal to the previously proposed speaker-dependent masker. However, female speech was not as effectively masked, and needs further investigation.

Since visual navigation systems using smart phones show information on small screens, user attention is likely to be distracted. Therefore, we are considering a portable navigation system using Augmented Audio Realty. However, if such equipment is used outdoors, very loud wind noise is recorded with the environmental sound. Thus, to reduce this wind noise, we use wind screen (ear muffs) and applied signal processing for wind noise reduction. In this paper, we compared a noise-controlled binaural earphone, and a bone conduction headphone. It was shown that Iterative Wiener filtering improves the speech intelligibility score dramatically with the former device.

In this paper, we compared two audio output devices for augmented audio reality applications. In these applications, we plan to use speech annotations on top of the actual ambient environment. Thus, it becomes essential that these audio output devices will be able to deliver intelligible speech annotation along with transparent delivery of the environmental auditory scene. Two candidate devices were compared. The first output was the bone-conduction headphones which can deliver speech by vibrating the skull, while normal hearing is left intact for surrounding noise since these headphones leave the ear canal open. The other is the binaural microphone/earphone combo, which is in a form factor similar to a regular earphone, but integrates a small microphone at the ear canal entry. The input from these microphones can be fed back to the earphone along with the annotation speech. In this paper, we compared the speech intelligibility of speech when competing babble noise is simultaneously given from the surrounding environment. It was found that the bone-conduction headphones can deliver speech at higher intelligibility than the binaural combo. However, with the binaural combo, we found that the ear canal transfer characteristics were altered significantly by closing the ear canal with the earphones. If we employed a compensation filter to account for this transfer function deviation, the resultant speech intelligibility was found to be higher than the bone-conduction headphones. In any case, both of these are found to be acceptable as audio output devices for augmented audio reality applications since both are able to deliver speech at high intelligibility even when significant amount of competing noise is present.

This chapter provides a general overview of audio data hiding. The general issues are discussed first, followed by the basic techniques used to hide data in audio signals, including bit stealing, spread spectrum methods, echo methods, and quantization index modulation. This is followed by a brief description of the recent proposals presented at the Institute of Electronics, Information, and Communication Engineers of Japan (IEICE) Multimedia Information Hiding (MIH) Technical Group Meetings.

This chapter proposes two data-hiding algorithms for stereo audio signals. The first algorithm embeds data into a stereo audio signal by adding data-dependent mutual delays to the host stereo audio signal. The second algorithm adds fixed delay echoes with polarities that are data dependent and amplitudes that are adjusted such that the interchannel correlation matches the original signal. The robustness and the quality of the data-embedded audio will be given and compared for both algorithms. Both algorithms were shown to be fairly robust against common distortions, such as added noise, audio coding, and sample rate conversion. The embedded audio quality was shown to be “fair” to “good” for the first algorithm and “good” to “excellent” for the second algorithm, depending on the input source.

When intelligibility is measured, objective estimation is more convenient than subjective. However, most objective estimation methods estimate monaural intelligibility using monaural signals. Therefore, it is necessary to estimate binaural intelligibility using binaural signals in order to take into account that a person listens using two ears. We evaluated an intelligibility estimation method using the Better Ear Model which selects better value out of the left and right feature values, and found that high estimation accuracy is possible. Accordingly, we proposed and evaluated an extension to this model which divides the signal into critical frequency bands, and takes the better value between left and right channel in each band (Band-Selection Better Ear Model). Furthermore, we tried estimation using machine learning in addition to regression analysis which we applied previously. Neural network and support vector machine were used here. Comparison between the Better Ear Model and the Band-Selection Bette...

Subjective speech intelligibility assessment is important for the development of outdoor public address system. However, as this assessment is difficult in many cases, we propose an objective speech intelligibility evaluation system that includes a machine learning technique. In this talk, we have proved a subjective evaluation and objective prediction of speech intelligibility using the outdoor public address systems at 10 locations in Sendai City, where impulse responses were recorded after the Great East Japan Earthquake. First, the results of the subjective intelligibility evaluation by different test word lists with the same sound field conditions showed that the root mean squared error (RMSE) was very small, not exceeding 7.0%. Next, we generated the intelligibility prediction model trained with true/false results of 22 subjects using the support vector machine (SVM). This prediction model extracted the feature vector, using the ITU-T P.563 speech quality feature set of the test speech signal. We ev...

This paper investigates methods to model inter-phrase or word context for continuous Japanese speech recognition. It is well known that in continuous speech, coarticulation between words or phrases induces allophonic variation of the beginning and ending phones in words or phrases. It was found that by compiling a network of contextdependent phonetic models which models these inter-word or inter-phrase context, recognition error reduction by 32 % can be achieved compared to models which do not account for inter-word context with task-dependent training, i.e. models that were trained with the same vocabulary as the test set. A more dramatic error reduction of up to 43% was possible with task-independent training. However, this will significantly increase the number of phonetic models required to model the vocabulary. With digit models, the increase in the number of models is 4 to 5 fold. To overcome this increase, we clustered the inter-word/phrase context into a few phonetic classes. Using one class for consonant inter-word context and two classes for vowel context, the recognition accuracy on digit string recognition was found to be virtually equal to the accuracy with unclustered models, while the number of phonetic models required was reduced by more than 50%.