https://hal.archives-ouvertes.fr/hal-00499180/document

Automatic speech recognition and speech variability: A review

Embodiments of the present invention are directed to a communication system that provides various automated operations, including searches of public and private resources for third party information and presentation of the located third party information by a mashup and preparation of a mashup to combine different communication modes or modes in a common session

Telephony discovery mashup and presence

A noise suppression device comprises subband SN ratio calculation means which receives a noise likeness signal, an input signal spectrum and a subband-based estimated noise spectrum, calculates the subband-based input signal average spectrum, calculates a subband-based mixture ratio of the subband-based estimated noise spectrum to the subband-based input signal average spectrum on the basis of the noise likeness signal, and calculates the subband-based SN ratio on the basis of the subband-based estimated noise spectrum, the subband-based input signal average spectrum and the mixture ratio.

Noise suppression device

Before executing a speech recognition, a composite acoustic model adapted to noise is generated by composition of a noise adaptive representative acoustic model generated by noise-adaptation of each representative acoustic model and difference models stored in advance in a storing section, respectively. Then, the noise and speaker adaptive acoustic model is generated by executing speaker-adaptation to the composite acoustic model with the feature vector series of uttered speech. The renewal difference model is generated by the difference between the noise and speaker adaptive acoustic model and the noise adaptive representative acoustic model, to replace the difference model stored in the storing section therewith. The speech recognition is performed by comparing the feature vector series of the uttered speech to be recognized with the composite acoustic model adapted to noise and speaker generated by the composition of the noise adaptive representative acoustic model and the renewal difference model.

Apparatus and method for speech recognition

Efficient speech recognition using subvector quantization and discrete-mixture HMMS

This paper proposes a new type of acoustic model called the discrete mixture HMM (DMHMM). As large scale speech databases have been constructed for speaker-independent HMMs, continuous mixture HMMs (CMHMMs) are needed to increase the number of mixture components in order to represent complex distributions. This leads to a high computational cost for calculating the output probabilities. The DMHMM represents the feature parameter space by using the mixtures of multivariate distributions in the same way as the diagonal covariance CMHMM. Instead of using Gaussian mixtures to represent the feature distributions in each dimension, the DMHMM uses the mixtures of the discrete distributions based on scalar quantization (SQ). Since the discrete distribution has a higher degree-of-freedom in terms of representation, the DMHMM is advantageous in representing the feature distributions efficiently with fewer mixture components. In isolated word recognition experiments for telephone speech, we have found that the DMHMM outperformed the CMHMMs when those models had the same number of mixture components.

Discrete mixture HMM

PROBLEM TO BE SOLVED: To lighten the burden of operator education by automatically scoring an answer made by an operator in charge of answering operation. SOLUTION: A voice feature quantity is detected (4, S202) from an input customer's voice signal, and matched with a feeling model set generated by modeling a plurality of predefined feelings in time series (16, S206) to generate feeling series by calls (10, S208). Respective feelings of a feeling series are converted into feeling points (22, S210) by using a feeling point list (28) in which the plurality of feelings and feeling points are made to correspond to each other. A value obtained by subtracting feeling points at the start of answering from feeling points at the end of the answering, an average value of feeling points of one call, etc., are calculated (26, S212) as answer evaluation points based upon the series of feeling points. COPYRIGHT: (C)2008,JPO&INPIT

Answer evaluating device and method thereof, and program and recording medium therefor

PROBLEM TO BE SOLVED: To improve the recognition precision and to precisely obtain a high rejection. SOLUTION: A feature parameter Xt is extracted from inputted voices (13). Then, candidates are defined in the order of the larger probabilities of a hidden Markov model(HMM) concerning the parameter Xt (33). Then, based on the distribution difference between the probability obtained by the category, which is the object by a phoneme, and the probability obtained by other category, the amount of variation, for example, the probability of an object category, is divided by the probability of an asymmetric category and which is beforehand made into a model in a statistical manner. Then, the probability (obtained by a reliability likelihood computing section 34) computed from the model is multiplied by the probability of the entire recognition candidates and accumulated for every path (32) and a network searching section 31 conducts a search work while being made to remain the candidates having higher likelihood.

Speech recognition method and its program recording medium

PROBLEM TO BE SOLVED: To provide a document retrieving device for extracting a keyword from a conversation, and for retrieving a document from the keyword, and for displaying the retrieval result, and for providing it to a user so that a user easily confirm the retrieval result. SOLUTION: This document retrieval device is provided with a title presenting and updating means for displaying a undisplayed document title from among retrieval results retrieved once even when any keyword is not extracted from a conversation, and configured to update the title display even when any keyword is not extracted. Thus, it is possible for a user to efficiently reach a target document by increasing the amounts of document information to be provided to the user. COPYRIGHT: (C)2008,JPO&INPIT

Document retrieval device

PROBLEM TO BE SOLVED: To divide voice data of a two-person conversation into a plurality of dialogue states. SOLUTION: An apparatus for dividing the dialogue state includes an utterance interval-detecting unit, a frame-extracting unit, an intra-frame utterance time rate-calculating unit, and a dialogue state-classifying unit. The utterance interval-detecting unit uses the voice data of two-person conversation as an input for detecting utterance intervals of each person. The frame-extracting unit arranges those utterance intervals in the order of elapsed time to output the predetermined number of utterance intervals as a single frame. The intra-frame utterance time rate-calculating unit calculates a time rate of the utterance interval for each frame. A frame-representative score-calculating unit determines the representative score, in a single frame in units of utterance intervals, based on the time rate of the utterance interval. A dialogue state-dividing unit divides the frame into at least three dialogue states, by comparing the time rate of the utterance interval with at least two thresholds. COPYRIGHT: (C)2011,JPO&INPIT

Takahashi Satoshi

Papers

Discrete mixture HMM

Answer evaluating device and method thereof, and program and recording medium therefor

Speech recognition method and its program recording medium

Document retrieval device

Apparatus, method and program for dividing dialogue state and recording medium