Mikio Mori

Recurrent neural networks for phoneme recognition

This paper deals with recurrent neural networks of multilayer perceptron type which are well-suited for speech recognition, specially for phoneme recognition. The ability of these networks has been investigated by phoneme recognition experiments using a number of Japanese words uttered by a native male speaker in a quiet environment. Results of the experiments show that recognition rates achieved with these networks are higher than those obtained with conventional non-recurrent neural networks.

Speech recognition based on a model of human auditory system

The paper deals with a new phoneme recognition system based on a model of human auditory system. This system is made up of a model of the human cochlea and a simple multilayer recurrent neural network which has feedback connections of self-loop type. The ability of this system has been investigated by a phoneme recognition experiment using a number of Japanese words uttered by a native male speaker. The result of the experiment shows that recognition accuracies achieved with this system in the presence of noise are higher than those obtained by a combination of frequency spectral analysis by DFT and a conventional feedforward neural network and that the cochlea model effectively prevents deterioration due to noise of recognition accuracy.

Kana-to-kanji conversion method using Markov chain model of words in bunsetsu

We previously proposed a kana-to-kanji conversion method of non-segmented kana sentences by using Markov chain model of words in sentence. However, we could not obtain the enough accuracy rate for conversion by this method. The cause is considered that the total number of the rules is not saturated in the dictionary of Markov chain probabilities of words in sentence. Therefore, we take notice that the total number of the rules is almost saturated in the dictionary of Markov chain probabilities of words in bunsetsu. In this paper, we propose a new kana-to-kanji conversion method by using this Markov chain model. That is, the new proposed method detects simultaneously the boundaries of kana bunsetsu in sentence and the boundaries of kana word in bunsetsu by using Markov chain model of kana words in bunsetsu, and then converts kana words to the candidates of kanji-kana word and selects the maximum likely candidate by using Markov chain model of kanji-kana words in bunsetsu. Through the experiments by using statistical data of daily Japanese newspaper, the previous proposed method (called Method-B1) and the new proposed method (called Method-B2) are evaluated by the criteria of the accuracy rate for conversion. From the results of the experiments, it is concluded that Method-B2 is superior to Method-B1 in the accuracy rate for conversion and is effective in kana-to-kanji conversion of non-segmented kana sentences.

Effects of wearing ear plugs in bone-conduction hearing via teeth

In this study, we measured the thresholds of bone-conduction hearing via teeth while wearing a mouthpiece with a bone-conductive actuator and ear plugs. It was found that the thresholds of bone-conduction hearing via teeth tend to be better than those of hearing with no ear plugs.

Word intelligibility of bone conductive sound when wearing ear plugs

In this study, we measured and compared the word intelligibility of air conductive and bone conductive sound when wearing ear plugs in both ears by speech audiometry. The correct responses of word intelligibility of bone conductive sound tended to be better than those of sound heard through air conduction even in a high signal-to-noise environment. This effect was only elicited when ear plugs were used in both ears.

Speech recognition based on subword units

Large vocabulary, isolated word recognition requires a large amount of training data proportional to the vocabulary size to characterize each individual word model. A subword‐unit‐based approach is a more viable alternative than the word‐based approach to overcome the problem of the training data size, since different words can share common segments in their representations in the former. This paper deals with a couple of isolated word recognition systems where the subword‐unit‐based approach is commonly employed, though their methods of segmentation are completely different. In one system a hidden Markov model is used to decompose a word into subword units (segments), and frequency spectra of those subword units are fed to a recurrent neural network to yield a subword code sequence for the word. This sequence is then recognized hopefully as the original word by a set of hidden Markov models for isolated words. In the other system subword boundaries within a word are detected by finding peaks of the delta...

Study on the principle of sound resonance in human whistling using physical models of a human vocal tract

The principle behind sound production in human whistling is relatively unknown. A good understanding of this principle would be beneficial to both trainer and trainee. Rayleigh identified whistling frequency as being determined by the mouth cavity and pointed out that earlier ideas that relate the sound-producing mechanism to the vibration of the lips are inaccurate. Wilson et al. reported that the human whistling resonant frequency is close to the Helmholtz resonant frequency, and through some physical-model-based experiments, they determined that the resonator can be excited by a flow through the smooth-edged orifices bounding the resonant cavity. Thus, the principles of whistling in terms of articulation have been reported to be based on the Helmholtz resonance. However, some whistlers can produce high-pitched sounds by blowing harder without changing the capacity of the resonance chamber, which is similar to a high-pitched sound produced by a wind instrument by air-column resonance. This work studies ...

A New Method to Reduce the Ambiguity of Japanese Phoneme Candidates Recognized by Recurrent Neural Networks

Up to now, the method to reduce the ambiguity of phoneme recognition using 2nd-order Markov chain model of phonemes, has been proposed and has been evaluated by phonem lattice simulated and limited to substitution error. However, the method will be necessary to demonstrate the effectiveness for the phoneme candidate lattice obtained by actual speech recognition devices. This paper deals with recurrent neural networks(RNN) which are well- suited for natural language processing of speech recognition, specially for phoneme recognition. The ability of these networks has been investigated by phoneme recognition experiments using a number of Japanese words uttered by a native male speaker in a quiet environment. A method to detect the locations of devoicing vowels using the short- time average energy has been also proposed, and evaluated. Form results of the experiments, it is shown that recognition rates achieved with RNN are higher than those obtained with conventional non-recurrent neural networks, and that the method to detect the locations of devoicing vowels is useful.

An Experimental System to Analyze or Synthesize Baby's Emotional Voice using the Varidation of Pitch Frequencies

It has been expected to utlize the technology of speech recognition and speech synthesis in order to improve the human interface. However, it is a difficult problem to extract and analyze emotional information contained in human voice in speech recognition and speech synthesis. This paper describes an experimental system to analyze and synthesize babys emotional voice using the perturbation parameters of pitch frequencies in order to improve human interface. The system is composed of three parts: the first part is to record and analyze the voice and to indicate the perturbation of pitch frequencies in real-time. The next part is to evaluate four perturbation parameters of pitch frequency for babys voice. Finally, babys voice is synthesized using 2-mass model of vocal cords vibration. From the experiment of several babys voices, it is shown that the proposed method is useful to analyze and synthesize babys emotional voice.

Networked Hearing Aids Using Wireless LAN Technologies

This paper proposes a novel scheme of hearing aid system which will release quite a lot of old people from light hearing loss or presbycusis, with its easy handling functions and high fidelity output. Hearing aids have been much improved over the past twenty years. However, they still require the troublesome fitting process, above all to old people. Most of hearing aid users point out the noisy performance in their reaction polls. The proposed system electronically connects hearing aids using both air and bone conducted voice signals each other via wireless local access, everywhere. The electronic transfer of voice signals enables to reduce the noise more effectively than the acoustic transfer, especially in noisy environments. It leads to global connections through high speed IP networks. Additionally, the amplifier for hearing loss compensation of the system can be controlled independently at each specified audiogram test frequency and gives less distortion than conventional sub-band flat gain controlled amplifiers. It also makes the hearing quality more natural. Thus, it is expected that isolating old people with hearing loss are able to evenly participate in the future ubiquitous network society as well as general people without hearing troubles. This paper describes the configuration, principles of the proposed scheme. The subjective performance test results show the possibility of outstanding quality improvement.