In a speech recognition system, the combination of a log-linear model with a multitude of speech features is provided to recognize unknown speech utterances. The speech recognition system models the posterior probability of linguistic units relevant to speech recognition using a log-linear model. The posterior model captures the probability of the linguistic unit given the observed speech features and the parameters of the posterior model. The posterior model may be determined using the probability of the word sequence hypotheses given a multitude of speech features. Log-linear models are used with features derived from sparse or incomplete data. The speech features that are utilized may include asynchronous, overlapping, and statistically non-independent speech features. Not all features used in training need to appear in testing/recognition.

Speech recognition utilizing multitude of speech features

PROBLEM TO BE SOLVED: To provide a speech recognition system and a speech processing system that can correctly accept a mode of expression that a user speaks like a conversations between persons. SOLUTION: The speech recognition system is equipped with an acoustic processing unit 11 which converts spoken speech data into a feature quantity, a speech section decision unit 12 which decides a speech section from the spoken speech data, a priority decision unit 18 which selects a speech section to be given priority out of speech sections decided by the speech section decision unit 12 based upon a specified priority acceptance criterion, and a matching unit 13 which calculates the degree of matching with a recognition vocabulary by using the feature quantity of the speech section selected by the priority decision unit 18 and a sound model. The priority decision unit 18 uses at least one of (1) speech section length, (2) power or SN of the speech section, and (3) time-series front/rear relation of the speech section as the specified priority acceptance criteria. COPYRIGHT: (C)2007,JPO&INPIT

Speech recognition system and speech processing system

The present invention relates to enhancing a quality of speech wherein speech quality degradation is reduced by removing noise from an unvoiced speech. The present invention comprises dividing an input speech into a voiced speech and an unvoiced speech, performing adaptive filtering on the voiced speech to remove a noise of the voiced speech, and performing special subtraction on the unvoiced speech.

Method of filtering speech signals to enhance quality of speech and apparatus thereof

PROBLEM TO BE SOLVED: To provide a speech processor capable of evaluating an impression of a speech of a first user.SOLUTION: The speech processor includes: a detection section that detects a first speaking period included in a first speech of a first user and a second speaking period included in a second speech of a second user; and a regulation part that regulates an overlap period in which the first speaking period and the second speaking period overlap with each other. The speech processor also includes; a calculation section that calculates a first speaking continuing period from a start point of the overlap period to an end point of the first speaking period; and an evaluation section that evaluates an impression of the first speech based on a piece of information relevant to a length of at least the first speaking continuing period.SELECTED DRAWING: Figure 1

Speech processor, speech processing method and speech processing program

A sound abnormality detection system for power device comprises an audio detection subsystem and a computer subsystem wirelessly communicated with the audio detection subsystem. The audio detection subsystem comprises an audio collection module, a video service module, an audio training and identification module, a storage module, and a first communication module. The audio collection module is connected with the video service module, the audio training and identification module and the first communication module in turn. The storage module is connected with the audio training and identification module bidirectionally. The computer subsystem comprises a host machine, an audio data display module, an alarm module, and a second communication module. The host machine is connected with the audio data display module, the alarm module, and the second communication module respectively. The detection system is flexible, low cost and highly reliable.

Sound abnormality detection system for power device and method thereof

PROBLEM TO BE SOLVED: To reduce a processing amount and deterioration in recognition precision by investigating an hourly change of a characteristic vector, substituting a value of before one point of time for a probability calculation value when a varying amount is less, further structuring distribution of probability to a binary tree structure and performing calculation of probability based on binary tree retrieval. SOLUTION: A characteristic amount hourly change detection means 31 receives the characteristic vector at one point of time from a voice analysis means 2 to compare it with the characteristic vector before one point of time. When the hourly varying amount of the characteristic vector is less than a prescribed threshold value as the comparison result, the means 31 issues instructions to a representative distribution probability calculation means 32 so as to omit processing. Further, the number of times of probability calculation are reduced remarkably by using the representative distribution of the binary tree structure. At this time, after arriving at a final hierarchy by the binary tree retrieval, the probability calculation is performed again related to the vicinity distribution of the arrived representative distribution, and a real probability calculation value instead of an approximate value is used. Thus, the voice recognition with less processing amount becomes possible while keeping the recognition precision.

Amano Akio

Papers

Voice recognition device