JPS6073593A - Phoneme dictionary preparation system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field of the Invention] The present invention relates to a phoneme recognition device, and in particular, when creating a phoneme dictionary, automatically attaching and checking sound 1N symbols to feature vectors.

For example, the present invention relates to a phoneme dictionary creation method that allows easy extraction of feature vectors of stationary phonemes.

[Prior art and problems]

When performing speech recognition, known speech is sampled in milliseconds, for example, and a feature vector is created by calculating the frequency spectrum.

A dictionary is created by adding labels indicating phonemes to these feature vectors.

For example, the problem arises when creating a phoneme dictionary that registers phonemes such as vowels and consonants.

The phonological nature of feature vectors at each time has traditionally been identified by humans visually observing the feature vectors.

The process involved identifying the phoneme category and labeling the input phoneme, which was very complicated.

In addition, in the case of stationary phonemes such as vowels and S that have the same frequency spectrum for a certain period of time, it is possible to automatically label them with feature vectors, but in this case, the boundaries between phonemes are excluded. This method automatically extracts only parts with extremely high stationarity and identifies only those parts, so there is a limit to the extraction range of feature vectors.

In any case, the relationship of conventional feature vector labeling to phonological recognition is not clear, and there is no guarantee of its effectiveness for phonological recognition. I couldn't decide what it was.

[Purpose of the invention]

The purpose of the present invention is to improve the above-mentioned drawbacks by identifying phonology using a simple method, and then inputting speech and performing this check using a method similar to phonological recognition to eliminate identification errors. Automatically identify stationary phonemes and create a phonological dictionary.

and a phonological dictionary that guarantees its effectiveness during phonological recognition.

The objective is to provide a method for creating documents.

[Structure of the invention]

To achieve this objective, the phonetic dictionary creation method of the present invention includes a feature vector time series converter stage for converting input speech data into a time series of feature vectors, and a speech segment extraction means for cutting out speech segments from the input speech data. A phonological symbol adding means for associating a phonological symbol string representing utterance content with audio data and adding a phonological symbol to the feature vector; a phonological dictionary storing the feature vector to which the phonological symbol has been added; a checking means for extracting a feature vector similar to the transmitted feature vector from a phonetic dictionary to detect whether or not the phonetic symbols match;
First, phonetic symbols are attached to the feature vectors at each time of the given input speech data, and these are stored in a phonological dictionary. First, they are similar to the feature vectors at each time given from other known input speech data. The present invention is characterized in that a feature vector with a phoneme symbol is extracted from the phoneme dictionary, and if the phoneme characteristics do not match, the feature vector data with a phoneme symbol is removed from the phoneme dictionary.

[Embodiments of the invention]

The configuration of an embodiment of the present invention will be explained based on FIGS. 1 and 2. FIG.

FIG. 1 is a configuration diagram of an embodiment of the present invention, and FIG. 2 is an explanatory diagram of its operation.

In the figure, l is a speech input section, 2 is an analysis section, 3 is a speech segment extraction section, 4 is a feature vector time series conversion section, 5 is a phoneme symbol string input section, 6 is a speech division section, and 7 is a phoneme segmentation section. 8 is a symbol adding section, 8 is a phonetic dictionary, 9 is a checking section, and 10 is a deletion section.

The speech input section 1 is to which a speech input signal of known phonemes is input.

The analysis unit 2 converts this audio input signal into, for example, the number 71! , 5' to measure the noise strength, send the sampling data from the feature vector time series to the converter 4, output the sampling time signal to the audio divider 6, and perform various analyzes. be.

The voice section extraction section 3 is the analysis section 2. The audio section is cut out based on the power information output from the .

The feature vector time series converting unit 4 performs, for example, Fourier transform on the sampling data transmitted from the analysis unit 2 and extracts its frequency components to obtain a time series of feature vectors.

The phoneme symbol string input section 5 is for inputting the phoneme name of the voice input input to the voice input section 1, and this phoneme name is input from, for example, a keyboard.

- The voice analysis section 6 divides the speech section cut out by the speech section extraction section 3 into phoneme units based on the phoneme symbols inputted from the phoneme symbol string input section 5, and adds phoneme symbols to the phoneme units.

The phoneme symbol assigning unit 7 assigns phoneme symbols of divided phonemes transmitted from the speech dividing unit 6 to the time series of the feature vectors transmitted from the feature vector time series converting unit 4.

The phoneme dictionary 8 stores the feature vectors with phoneme symbols sent from the phoneme symbol adding section 7.

The checking unit 9 checks whether the feature vectors stored in the phonetic dictionary 8 and the phonetic symbols attached to them are accurate. .

A newly transmitted known phoneme feature vector and a feature vector that is highly similar to this feature vector are extracted from the phoneme dictionary 8, these phoneme symbols are compared, and if they do not match, a mismatch signal is output. Control is performed to remove this feature vector stored in the phoneme dictionary 8 of the lever.

The deletion unit 10 deletes the feature vector from the phoneme dictionary 8 based on the discrepancy signal output from the check unit 9, and performs control such as removing phoneme symbols, for example.

Next, the operation of the present invention will be explained with reference to FIGS. 1 and 2.

(Li) Creation of a phoneme dictionary First, a phoneme dictionary is created. In this case, switches St and S2 are both switched to the phoneme symbol adding unit 7 side. Then, the voice signal strength of the known phoneme (input to the voice input unit The feature vector time series converter 4
is converted into a time series of feature vectors. On the other hand, the power information obtained from the analysis section 2 is sent to the speech section cutting section 3 and cut out into speech sections. At this time, since a phoneme symbol string is input separately from the phoneme symbol string input section 5, this speech segment is divided into phoneme units by the speech division section 6, and phoneme symbols are assigned. The feature vector created by the feature vector time series converter 4 is sent to the phonetic symbol assigning section 7, but at this time, the speech dividing section 6 sends the segmented data of the phoneme unit and its phonetic symbol. A phoneme symbol is assigned to each time-series feature vector, and this is stored in the phoneme dictionary 8.

In this way, a phonetic dictionary is created.

(2) Checking the phoneme dictionary When checking the phoneme dictionary 8 created as described above, switch the switches Sl and S2 to the checking unit 9 side.Then, the audio signal of the known phoneme is sent to the audio input unit 1.
The phonetic symbol string of t is input to the phonetic symbol string input section 5.
is input. At this time, the time series of feature vectors output from the feature vector time series converter 4 is transmitted to the checker 9.

The checking unit 9 extracts from the phonetic dictionary 8 those that are highly similar to this feature vector F.
Since the divided data of phoneme units and phoneme symbols outputted from the speech dividing section 6 are transmitted to.

It is checked whether the phonetic symbol of the feature vector extracted from the phonetic dictionary 8 matches the phonetic symbol transmitted from the speech dividing section 6. When these phoneme symbols do not match, the checking section 9 outputs a mismatch signal, and the deletion section 1O deletes the feature vector stored in the phoneme dictionary 8.

For example, the feature vector f=・f of the phoneme S as shown in FIG.
2... When frn is compared with the phonetic dictionary 8 and the feature vector △S++△02...△Sm is extracted, the feature vector △Ct with the phonetic symbol C will be deleted from the phonetic dictionary 8. . In this way, by repeating checks using a plurality of known phonemes, inaccurate features are removed from the phoneme dictionary 8, and accurate ones are left, which can be used to identify correct phonemes. It can be performed.

〔Effect of the invention〕

According to the present invention, it is not only possible to automatically and easily identify phonemes, but also to eliminate errors caused by inaccuracies by checking after the phoneme dictionary is created.

Since only accurate feature vectors can be stored,
During recognition, an effective phonetic dictionary with very high recognition accuracy can be obtained. Furthermore, depending on the speech segmentation method, phoneme identification can be performed to the very limits of its performance.

[Brief explanation of drawings]

FIG. 1 is a configuration diagram of an embodiment of the present invention, and FIG. 1A-2 is an explanatory diagram of its operation. In the figure, 1 is a speech input section, 2 is an analysis section, 3 is a speech segment extraction section, 4 is a feature vector time series conversion section, 5 is a phoneme symbol string input section, 6 is a speech division section, and 7 is a phoneme symbol addition section. section, 8 is a phoneme dictionary, 9 is a check section, and IQ is a deletion section.

Claims (1)

[Claims] Feature vector time series conversion means converts input speech data into a time series of feature vectors; speech section extraction means extracts speech sections from the input speech data; A phonetic symbol adding means for adding a phonetic symbol to the feature vector; a phonetic dictionary that stores the feature vector to which the phonetic symbol is added; It has a checking means for extracting phonetic symbols from a dictionary and detecting whether or not they match. First, a phonetic symbol is attached to each quadruple feature vector of given input speech data, and these are stored in a phonetic dictionary. First, a feature vector with a phoneme symbol that is similar to the feature vector at each time given from other known input speech data is extracted from the phoneme dictionary, and if the phoneme characteristics do not match, the phoneme symbol is extracted from the phoneme dictionary. A phonological dictionary creation method characterized by removing feature vector data with a tagged feature from a phonological dictionary.