Sharp Kabushiki Kaisha Corporation

Shin Kamiya

Toru Ueda

Fumio Togawa

Atsuo Tanaka

A user system incorporating a voice instruction recognition processing portion and a transmission control portion is connected through a communication line to a center system incorporating a reception control portion, a voice instruction recognition processing portion and an output control portion. Processing for obtaining a voice instruction content based on a waveform analysis result of the voice instruction is shared by voice instruction recognition processing portions. This share prevents decrease of a recognition rate occurred in the case of disposing a voice analysis recognition processing function only on the center system. The share also eliminates waiting time for transfer of lexical grammar data generated in the case of disposing the voice analysis recognition processing function only on the user system. Further, a processing load of the user system is decreased. Thus, high recognition capability is obtained even with voice input from a user terminal low in power consumption.

Speech input communication system, user terminal and center system

This dynamic programming includes a learning mode and an evaluation mode. In the learning mode, an accumulated DP path matrix is determined by a DP matching among a plurality of learning pattern vectors belonging to one category and a plurality of learning pattern vectors belonging to a similar category similar to the foregoing category and a standard pattern vector belonging to the foregoing category. Then path weights are calculated depending on the values of each element of the accumulated DP path matrix. In the evaluation mode, for performing a DP matching between evaluation pattern vectors and the standard pattern vector, a distance between elements of the above two pattern vectors is calculated with weighting by the above path weight. Thus, the optimum path in the evaluation mode is modified toward a frequent optimum path composed of a series of elements having high values in the foregoing accumulated DP path matrix.

Learning dynamic programming

A Lombard voice recognition method for recognizing a voice input in a noisy background includes a step of matching and a step of warping. The step of matching matches a frequency spectrum of a feature pattern derived from the input voice to a frequency spectrum of a standard pattern. The step of warping warps a lower frequency spectrum of the feature pattern than a lower frequency side with respect to a frequency axis in case that a noise level of the background in higher than a predetermined noise level.

Lombard voice recognition method and apparatus for recognizing voices in

A learning method of a neural network, in which from a set of learning patterns belonging to one category, specific learning patterns located at a region close to learning patterns belonging to another category are selected and learning of the neural network is performed by using the specific learning patterns so as to discriminate the categories from each other.

Learning method of neural network

The past record of the synaptic weight values set in the learning of a neural network is stored in a weight record memory. The past stored in the weight record memory is supplied to a control unit. If there exists a synaptic connection representing a record of weight values which have been used in a predetermined number of learning processes just prior to the present learning process and which satisfy a predetermined condition, the synaptic weight value used in the succeeding learning processes for the synaptic connection is re-set to a predetermined value by a weight setting unit. That is, the past record of the synaptic weight values is monitored, and the synaptic weight value which has been set in a learning process can be re-set as required.

Apparatus for changing individual weight value of corresponding synaptic

A control device for controlling the learning of a neural netowrk includes a monitor for monitoring weight values of synapse connections between units of the neural netowrk during learning of the neural network so as to update these weight values. When one of the weight values satisfies a preset condition, the weight value is updated to a predetermined value such that configuration of the neural network is determined in an optimum manner.

Control device for controlling learning of a neural network

A pattern recognition device has a DP matching section. The DP matching section performs frequency expansion DP matching to a standard pattern and a characteristic pattern obtained from input voice waveform to obtain a DP score and DP path pattern. It is determined by means of a category identification neural network using the DP path pattern obtained from the DP matching section whether a category of the standard pattern and a category of the characteristic pattern are the same, and a determination result corresponding to the degree of identification is obtained. A normalized DP score, which is the DP score normalized for individual differences within a required range, is then obtained in a divider by compensating the DP score using the determination result.

Pattern recognition device using a neural network

A subject associating device which includes concept table storage for accommodating a concept table; subject pattern storage for accommodating subject patterns representative of subjects of sentences; a subject pattern generator for converting a sequence of learning phrases or words into a sequence of independent word concepts for generating histograms for the independent word concepts and for storing a pattern of the histograms in the subject pattern storage as a subject pattern; and a subject associating unit for converting the sequence of phrases or words input into the sequence of independent word concepts, for generating the histogram for the independent word concept, for performing a matching between the pattern of independent word concepts of the input sentence and the subject pattern accommodated in the subject pattern storage, and for selecting the subject pattern, which has the small matching distance, as the subject pattern descriptive of the subject of the input sentence.

Subject and word associating devices

Syllables are recognized in voice data obtained from an input voice signal. Character arrays are used to represent the types and time frames of syllables with which the voice data are compared.

Syllable recognition system

A word processor stores sequences of key operations corresponding to its many word processing functions and monitors the sequence of key operations by the user. If it is found that the user has operated keys in one of the stored sequences for a specified number of times, sentences explaining the corresponding function of the word processor are displayed.

Word processor displaying instructions by monitoring user's key

An inputted sound signal is sampled at intervals over a period and cepstrum coefficients are calculated from the sampled values. Cepstrum sum, distance and/or power are calculated and compared with appropriately preselected threshold values to distinguish voice (vowel) intervals and noise intervals. The ratio of the length of the voice intervals to the sampling period is considered to determine whether the sampled inputted sound signal represents voice or noise.

Growing community of inventors

Nara, Japan

Shin Kamiya