JP2002116788A - Speech recognizing device and method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a speech recognizing device and method capable of improving the recognition accuracy in phonemic boundaries. SOLUTION: The labels of the phoneme in which the phonemic pieces for standard patterns are included and the labels indicating the positions of the phonemic pieces are added to the characteristics of these phonemic pieces and are previously stored in an HDD 40. A CPU 10 integrates the labels of the same plural phonemes continuous among the labels obtained as the result of the recognition of the phonemic pieces, thereby providing the result of the recognition of the phonemes.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a phoneme recognition apparatus and method for performing phoneme recognition using recognition results of phoneme segments.

[0002]

2. Description of the Related Art Conventionally, acoustic characteristics of speech are obtained by phoneme segment length (length of a plurality of phoneme segments obtained by decomposing one phoneme), and identification information indicating the feature of phoneme segment length and the contents of the phoneme is obtained. (Generally called a label) are prepared in advance. When speech recognition is performed, a feature of a phoneme length obtained from a speech to be recognized is compared with a feature (a so-called standard pattern) of a labeled phoneme prepared in advance. The value of the distance between the two features is used for this comparison. The feature of the recognition target phoneme and the features of the plurality of standard patterns are compared, and the label of the standard pattern with the shortest distance is determined as the speech recognition result. Further, a speech recognition method for correcting a phoneme recognition result based on a statistical appearance probability has also been proposed.

[0003] For example, as shown in Table 1 below, when learning a speech “Kai-satsu”, the acoustic features of the speech are decomposed in units of phonemes, and [k], [k],
Labels such as “a”, “a”,.

If there is no standard pattern whose distance calculation result is within an allowable range for the characteristics of a phoneme segment extracted from the speech to be recognized, the speech result is treated as being unrecognizable (see Table 1). 1 and?).

[0005]

[Table 1]

[0006]

The phonemes labeled as one type by the conventional method also tend to have a large variance when observing their characteristics, but have a large variance, and are likely to be erroneously recognized at phoneme boundaries.

An object of the present invention is to provide a phoneme recognition apparatus and method capable of improving the recognition accuracy of a phoneme boundary portion.

[0008]

In order to achieve the above object, according to the first aspect of the present invention, a speech to be recognized is speech-recognized in phoneme units, and phoneme recognition is performed using the speech recognition result. In the phoneme recognition device to be performed, a first feature per phoneme unit extracted from a speech whose utterance content is known in advance,
A storage unit that stores a plurality of sets of labels indicating the positions of the features and labels of phonemes including the phoneme segments in association with each other, and a feature extraction unit that extracts a second feature in speech unit units from the speech to be recognized And a phoneme segment recognizing means for detecting a first feature most similar to the second feature extracted by the feature extracting means from a plurality of sets of first features stored in the storage means; Label extracting means for extracting a label of a phoneme associated with the extracted first feature, label integrating means for integrating the same continuous label in the extracted label of the phoneme into one label, Output means for outputting the generated label as a phoneme recognition result.

According to a second aspect of the present invention, in the phoneme recognition device according to the first aspect, a first feature is extracted from the learning target speech by the feature extracting means, and further, a position to be associated with the first feature is extracted. And input means for inputting a label indicating phoneme and a phoneme label.

According to a third aspect of the present invention, in the phoneme recognition apparatus according to the first aspect, the phoneme is divided into three parts, a front part, a center part, and a rear part, and a phoneme fragment included in the divided part is included. In contrast, a label indicating a position of any of the three parts is associated with the first feature.

According to a fourth aspect of the present invention, there is provided a phoneme recognition method for performing speech recognition of a speech to be recognized in units of phonemes and performing phoneme recognition using the speech recognition result. A plurality of sets of the extracted first feature of each phoneme unit, a label indicating the position of the feature, and a label of a phoneme including the phoneme segment are associated with each other and stored in the storage unit. A second feature is extracted for each phoneme unit, and a first feature that is most similar to the second feature extracted by the feature extracting unit is selected from a plurality of sets of first features stored in the storage unit. Detecting, extracting a label of a phoneme associated with the detected first feature, integrating successive identical labels in the extracted label of the phoneme into one label, and As a phoneme recognition result Characterized in that comprises a output.

According to a fifth aspect of the present invention, there is provided the phoneme recognition method according to the fourth aspect, wherein a first feature is extracted from a speech to be learned, and a label indicating a position and a phoneme to be associated with the first feature. Is input, and the input label and the feature extracted from the learning target speech are stored in the storage unit in association with each other.

According to a sixth aspect of the present invention, in the phoneme recognition method according to the fourth aspect, the phoneme is divided into three parts, a front part, a center part, and a rear part, and a phoneme fragment included in the divided part is included. In contrast, a label indicating a position of any of the three parts is associated with the first feature.

[0014]

Embodiments of the present invention will be described below in detail with reference to the drawings.

First, a speech recognition method to which the present invention is applied will be described.

In the phoneme recognition method according to the present embodiment, a phoneme labeled as one type is classified according to its feature and a feature is generated again to obtain a finer feature.
The recognition rate is improved by recognizing the same type of labeled phonemes as the same phoneme.

For example, the phoneme “A” is replaced by “ha”, “m”
The phonemes "a", "ta" and "i" are "hi" and "mi"
A plurality of phoneme words are learned by performing a plurality of labeling on one phoneme as a front part, a middle part, and a rear part of the appearance position of “ti” and the sound. Here, h is a label meaning a front part in a phoneme. m is a label meaning a central part in a phoneme. t is a label indicating the rear part of the phoneme. After these labels, a label indicating the content of the phoneme, in this case a is added.

Using such a label notation method, learning for attaching the above-mentioned label to the acoustic feature of the phoneme segment is performed in advance.

When speech recognition is performed using such learning results (characteristics of labeled speech segments), the following results in Table 2 are obtained.

[0020]

[Table 2]

In the present embodiment, a phoneme segment to be recognized first, in the above example, for example, a phoneme segment corresponding to "k" is compared with a plurality of sets of standard patterns, and the most similar (smallest distance) is obtained. A label, in this case, a label string “hk” indicating the position of the phoneme segment and the phoneme content is obtained. Hereinafter, a label string “ma” on the standard pattern side most similar to the next phoneme segment to be recognized is obtained.

As described above, in the label sequence arranged in time series in which the speech recognition processing is performed on the speech to be recognized to obtain the phoneme labels including the positions of the phoneme segments, the label indicating the phoneme content is The same and continuous ones are combined into one. In the example of Table 2 above, since “k” is common and continuous between “hk” and “mk”, “k” is finally determined as a phoneme recognition result including the phoneme segment. When a phoneme label is obtained in this way, four frames or 12
Even if “ma” is erroneously recognized as “ta” as in the frame, the recognition result can be processed as “a” because the phonemes are in the same category of “a”.

Next, a phoneme recognition apparatus using the above-described phoneme recognition method will be described. FIG. 1 shows a circuit configuration of the phoneme recognition device of the present embodiment.

Since a general-purpose computer such as a personal computer can be used as the phoneme recognition device, the description of the hardware configuration will be briefly described.

In FIG. 1, a CPU 10, a system memory 20, an input / output interface (I / O) 30, a hard disk (HDD) 40, an input device 50, and a display device 6
0 and the like are bus-connected. The CPU 10 performs the phoneme recognition according to the present invention using a phoneme recognition program described later stored in the hard disk 40. System memory 20
Temporarily stores various data related to phoneme recognition. I
A / O30 inputs a voice to be learned or a voice to be recognized from the microphone.

The hard disk 40 stores a phoneme recognition program, an operating system (so-called OS) for controlling the entire system, and the like. The input device 50 has a mouse and a keyboard, and inputs information. The display device 60 displays information input from the input device 50, a phoneme recognition result, and the like.

The operation of the phoneme recognition device will be described with reference to FIGS. FIG. 2 shows the contents of a program relating to phoneme learning. FIG. 3 shows the contents of a program relating to phoneme recognition.

(Learning process) The voice (signal) input from the microphone is temporarily stored in the system memory 20 and then the characteristics of the sound are analyzed in the same manner as in the prior art. It is temporarily stored (step S100 → S110 in FIG. 2).

The results of the analysis are displayed on the display device 40 in chronological order. The user inputs a label corresponding to each acoustic analysis result from the keyboard of the input device 50. This label is the label described in the learning column of Table 2 (more precisely, a level including a label indicating the position of a phoneme and a label indicating the content of a phoneme including the phoneme). The CPU 10 stores the input analysis result of the label and the phoneme segment as a set of records in the hard disk 40. The above processing is performed on a plurality of phoneme segments extracted from the input speech (steps S120 → S130 in FIG. 2).

(Phone element recognition) In the case of phoneme recognition, a voice to be recognized is input from a microphone. The input voice is temporarily stored in the system memory (step S200 in FIG. 3). The input speech is subjected to acoustic analysis in the same manner as in the past in units of phonemes (step S210 in FIG. 3).

The results of the acoustic analysis are compared with labeled features (standard patterns) stored in the hard disk 40, and the most similar labeled features are detected. Labels are extracted from the detected labeled features and stored in the system memory 20 in chronological order ("Recognition" in Table 2).
Column label). The levels arranged in time series will be referred to as label series. The CPU 10 uses the program of FIG. 4 to be described later to perform a process of integrating consecutive phoneme labels having the same phoneme label into one label (step S220 in FIG. 3). The integrated label sequence is displayed on the display device 60 as a phoneme recognition result,
It is printed out by a printer (not shown) (see the label in the column of "final recognition" in Table 2; step S230 in FIG. 3).

(Integration of Labels) One processing method for detecting a portion having the same phoneme label continuously from the label sequence will be described with reference to FIG.

When the program shifts to the program of FIG. 4, an initial setting process is first performed. Here, a temporary immediately preceding label to be compared with the first label is set (step S3 in FIG. 4).
00). This label uses a label symbol that cannot actually exist.

Next, the first label in the label sequence obtained by the above-described recognition processing, that is, the label of the first frame, "hk" in the example of Table 2, is read out from the system memory 20, and the phoneme content is read out of it. Is extracted (step S310 in FIG. 4). The extracted “k” is compared with the initially set temporary immediately preceding label. Here, since a determination of mismatch is obtained, the label “k” to be compared is temporarily stored in the system memory 20 as a label to be integrated (step S320 → S in FIG. 4).
330).

Since all the labels have not been processed, the procedure is as follows: steps S330 → S340 → S31 in FIG.
After that, the recognition result of the second frame, that is, “tk” in Table 2 is read from the system memory 20, and the phoneme level “k” is extracted. This label "k" and the label "k" extracted from the label of one frame last time
Are compared (step S320 in FIG. 4). In this determination, a YES (match) determination is obtained, and the procedure returns to steps S320 → S340 → S310. As described above, when the phoneme label at a certain time is extracted from the label sequence of the recognition result, the same phoneme label is continuously detected by comparing with the phoneme label extracted at the previous time (one frame before). . In the case where the label does not match the label at the previous time, for example, in the case of the phoneme label of the fifth frame and the phoneme label of the sixth frame in Table 2, it is determined that the label at the current time is an utterance of a phoneme piece of another label. , The label name is temporarily stored in the system memory 20 (FIG. 4).
Step S330).

If the previous label and the current label match, the current label is not stored in the system memory 20. Therefore, a continuous portion of the same label name exists in the label sequence of the recognition result. In this case, in this processing procedure, the head label of the continuous part is detected and stored in the system memory 20, so that a plurality of consecutive same phoneme labels are integrated into one phoneme label.

When the above processing is performed on all the label sequences, the label integration results "ka", "i", "sa",...
・ ・ Is obtained.

The following embodiment can be carried out in addition to the above-described embodiment.

1) In the above-described embodiment, a learning function, that is, a feature of a speech whose utterance content is known in advance is extracted, and a label indicating the position and the content of the extracted feature is given to the hard disk storage device. It has the function of storing the attached features. However, the learning function and the phoneme recognition function may be divided. In this case, externally labeled feature data is provided to the phoneme recognition device. As an input method, various methods such as communication and a method using a portable recording medium can be used.

2) A program for extracting acoustic features during learning and a program for extracting acoustic features during phoneme recognition can be shared.

[0041]

As described above, according to the present invention, even if a phoneme is recognized by a phoneme piece and is erroneously recognized as a phoneme piece at another position, the label of the phoneme containing the phoneme piece is changed. Performing correct phoneme recognition by integrating the same continuous part of the phoneme segment recognition result (the phoneme label of the first feature most similar to the second feature) because it is associated with the first feature. Can be. As a result, a voice portion that spans between phonemes such as a cross section or
It is called a flat section, and it is possible to output a phoneme that is closer to a part having a different feature within the same phoneme as a phoneme recognition result.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a circuit configuration according to an embodiment of the present invention.

FIG. 2 is a flowchart showing the contents of a learning program according to the embodiment of the present invention.

FIG. 3 is a flowchart showing the contents of a phoneme recognition processing program according to the embodiment of the present invention.

FIG. 4 is a flowchart illustrating details of a phoneme recognition process according to the embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 10 CPU 20 System memory 30 I / O 40 HDD 50 Input device 60 Display device

Claims (6)

[Claims] 1. A phoneme recognition apparatus for performing speech recognition of a speech to be recognized in phoneme units and performing phoneme recognition using the speech recognition result, wherein a phoneme unit extracted from speech whose utterance content is known in advance is provided. A plurality of sets of the first feature, a label indicating the position of the feature, and a label of a phoneme including the phoneme segment in association with each other, and stored. A feature extracting unit for extracting a feature; and a phoneme detecting a first feature most similar to the second feature extracted by the feature extracting unit from a plurality of sets of the first features stored in the storage unit. Piece recognition means; label extraction means for extracting a label of a phoneme associated with the detected first feature; label for integrating the same consecutive labels in the extracted phoneme labels into one label And integrating means, phoneme recognition apparatus characterized by the integrated label comprises an output means for outputting a phoneme recognition result. 2. The phoneme recognition device according to claim 1, wherein the feature extracting means extracts a first feature from the learning target speech, and further associates the first feature with the first feature.
A phoneme recognition device comprising input means for inputting a label indicating a position and a phoneme label. 3. The phoneme recognition device according to claim 1, wherein the phoneme is divided into three parts: a front part, a center part, and a rear part.
A label indicating the position of any of the three parts is assigned to the first segment of the phoneme included in the divided part.
A phoneme recognition device characterized by associating with a feature of a phoneme. 4. A phoneme recognition method for recognizing a speech to be recognized in phoneme units and performing phoneme recognition using the speech recognition result, wherein a phoneme unit extracted from speech whose utterance content is known in advance is provided. A plurality of sets of the first feature, a label indicating the position of the feature, and a label of a phoneme including the phoneme segment are associated with each other, and stored in the storage unit. And extracting a first feature most similar to the second feature extracted by the feature extracting means from a plurality of sets of first features stored in the storage means. Extract the label of the phoneme associated with the extracted first feature, integrate the same continuous label in the extracted label of the phoneme into one label, and output the integrated label as a phoneme recognition result To do Phoneme recognition method characterized by the following. 5. The phoneme recognition method according to claim 4, wherein a first feature is extracted from the speech to be learned, and a label indicating a position and a phoneme label to be associated with the first feature are input. And storing the input label and the feature extracted from the learning target speech in the storage unit in association with each other. 6. The phoneme recognition method according to claim 4, wherein the phoneme is divided into three parts: a front part, a center part, and a rear part;
A label indicating the position of any of the three parts is assigned to the first segment of the phoneme included in the divided part.
A phoneme recognition method characterized by associating with a feature of a phoneme.