JPH0430599B2 - - Google Patents

Description

[Detailed description of the invention]

INDUSTRIAL APPLICATION FIELD The present invention relates to a speech recognition method that uses a plurality of standard patterns for each word, syllable, or phoneme. Structure of the conventional example and its problems In order to accommodate unspecified speakers, a speech recognition method that uses a plurality of standard patterns for each word, syllable, or phoneme is an effective method. In such a method, it is possible to increase the number of standard patterns in order to accommodate a larger number of speakers, but conversely, the large number of standard patterns increases the variation in the standard patterns, causing misrecognition. . Table 1 shows an example of the relationship between the number of standard patterns and the recognition rate when five vowels are used as standard patterns.

[Table] In this speech recognition method that uses multiple standard patterns, learning is performed to adapt to the speaker and select a standard pattern that is suitable for the speaker from among the multiple standard patterns. It is possible to reduce misrecognition and improve the recognition rate for the speaker. As a learning method aimed at adapting to such speakers, learning has conventionally been carried out for each word, syllable, or phoneme. However, in these conventional methods, in order to adapt all standard patterns to speakers, a certain amount of learning samples are required for each word, syllable, or phoneme. The problem was that it was slow. In addition, the samples used for learning must be highly reliable for each word, syllable, or phoneme, and in this sense, an operator (hereinafter referred to as a teacher) is required during learning, and automatic learning without supervision is I had a difficult problem. Purpose of the Invention The present invention has been made in view of the problems of the conventional methods as described above, and its purpose is to create a standard that is adapted to the speaker through learning in a speech recognition method that uses a plurality of standard patterns. A learning method that improves the recognition rate for the speaker by selecting patterns, is fast, highly efficient, does not rely heavily on the reliability of individual learning samples, and is therefore capable of automatic learning without supervision. Our goal is to provide the following. Structure of the Invention Word sounds, syllables, and phonemes uttered by a specific speaker are related to each other. For example, Figure 1 shows the first formants of the five vowels of two speakers A and B.
F ₁ and the position of the second formant F ₂ are shown. As shown in this figure, there are phonemes that are close between two speakers, such as |a| for speaker A and |o| for speaker B, but if you pay attention to each speaker Phonemes are located a sufficient distance apart. Furthermore, if we focus on a particular speaker, there is a certain positional relationship between phonemes, and once the position of one or several phonemes is determined, it is possible to limit the positions of the remaining phonemes to some extent. be. Based on the above-mentioned properties of speech, the present invention combines a plurality of individual, syllable, and phonetic parameter vectors for each speaker to create one parameter vector, and clusters the set of the combined parameter vectors. Then, the representative vector of each cluster is set as a set of standard patterns, and learning is performed for the speaker based on the set of standard patterns. Then, a set of standard patterns suitable for the speaker is selected, and the selected standard patterns are used to identify the voice. DESCRIPTION OF EMBODIMENTS An embodiment of the present invention will be described in detail below while comparing it with a conventional example. First, we will discuss the standard pattern creation method and learning method, and then we will discuss the recognition results during learning. () Conventional method Standard pattern creation method After clustering a set of parameter vectors for each phoneme category and dividing into an appropriate number of clusters, the representative vector of each cluster is used as the phoneme standard pattern for that phoneme category. Alternatively, a set of parameter vectors of all phoneme categories is clustered and divided into an appropriate number of clusters, and then the representative vector of each cluster is taken as the phoneme standard pattern of the phoneme category to which the cluster belongs. Learning method A standard pattern suitable for the speaker is selected for each phoneme based on the frequency of use of the standard pattern for each phoneme that the learning sample refers to during identification and the reliability of the learning sample during identification. At this time, it is preferable to reduce the number of learning samples to be selected as the number of learning samples increases. () Method according to the present invention Method for creating a set of standard patterns For each speaker, parameter vectors (number of dimensions n) of all phonemes (number of phonemes m) are combined to create one parameter vector (number of dimensions m×n). create. Second
The figure shows an example in which the phoneme is five vowels and the parameter vector is two-dimensional. Next, the set of combined parameter vectors is clustered, and the representative vector of each cluster is used as a set of phoneme standard patterns. In this clustering, weighting may be performed according to the importance of each type of phoneme, or weighting may be placed on vectors near phoneme boundaries that are likely to cause misidentification. Learning method: Based on the reliability of the training sample for identification and the frequency of use of the standard patterns that the training sample referred to during identification, the usage frequency is calculated based on the reliability of the entire set of standard patterns. Select a set of standard patterns that match the speaker. At this time, it is preferable to reduce the number of learning samples to be selected as the number of learning samples increases. In Figure 3, in a two-dimensional plane consisting of five vowels as phonemes and F ₁ and F ₂ as parameter vectors,
An example is shown in which four standard patterns are created for each phoneme. In the figure, the black circles indicate the selected standard patterns, and figure a shows the case where the conventional method is used, and b shows the case where the method of the present invention is used. Hereinafter, the recognition results during learning by the method of the present invention will be described in comparison with the conventional method. The recognition experiment was conducted on five vowels under the conditions described below. In other words, linear predictive analysis was performed on the voices uttered /ieaou/ five times each by 15 men in succession, and 600 frames (15 men x 5 vowels x 8 frames) from the two utterances were used to create a phonological standard pattern. As a recognition sample, 675 frames (15
After obtaining the LPC cepstral coefficients of (first name x 5 vowels x 9 frames), the projection matrix to Fisher space that maximizes the Fisher ratio is obtained using the sample for creating the phonological standard pattern, and the projection matrix for the standard pattern creation and recognition sample is We performed a recognition experiment using a sample that was projected onto a four-dimensional Fisher space. Figure 4a has 10 standard patterns for each phoneme. This figure shows the false recognition rate in the process of proceeding with learning using the two methods described above. In addition, since the false recognition rate before learning differs between the two methods, FIG. 4b shows the ratio with the false recognition rate before learning being 100. In FIG. 4, × indicates the result by the conventional method, and ○ indicates the result by the method of the present invention.
The figure shows the results for two different phoneme learning orders. That is, one is in the order of the number of identification errors (o→u→a→i→e), and the other is in the reverse order. In the conventional method, each time each phoneme is learned, one is selected from ten standard phoneme patterns belonging to that phoneme category, and in the method of the present invention, 10, 6, 4, 3, 2 are selected depending on the number of phonemes to be learned. , 1, and set, and finally both select one standard pattern for each phoneme. As can be seen from FIG. 4, in the conventional method, the reduction in the false recognition rate is greatly influenced by the order of phonemes to be learned. However, the method according to the present invention shows a nearly uniform decline, and the influence of the type of learned phoneme is small. Figure 5a shows how the false recognition rate decreases with respect to the number of training samples used until 1 standard pattern is finally selected for each phoneme, with 10 standard patterns for each phoneme. It is something. Figure b shows the rate of misrecognition before learning as 100. In FIG. 5, × indicates the result by the conventional method, and ○ indicates the result by the method of the present invention. As can be seen from the figure, compared to the conventional method, the method of the present invention shows a false recognition rate close to the final state with fewer learning samples, indicating the speed of adaptation. Table 2 shows, when one standard pattern is finally selected for each phoneme, whether the speaker belongs to the cluster that was selected as the representative vector when creating the standard pattern. . From this table, the method according to the present invention selects for all speakers a standard pattern that is a representative vector of the cluster to which each speaker belonged, whereas in the conventional method, a large proportion of different patterns are selected. This shows that the learning method according to the present invention is highly reliable.

[Table] Although the present invention has been described with reference to specific embodiments, the present invention is not limited to the above embodiments. For example, phonemes other than vowels may be used as the standard pattern, or syllables may be used instead of phonemes. , you can also use words. Effects of the Invention As described above, the present invention has the following effects. Adaptation to speakers is possible with a small number of learning samples (learning speed is fast). The misrecognition rate can be reduced without being greatly influenced by the type of standard pattern to which the learning sample belongs. Even if the reliability of the identification results of individual learning samples is not a concern, the reliability of the entire set is a concern, so the possibility of selecting an inappropriate standard pattern is small. That is, according to the present invention, when selecting a standard pattern suitable for a speaker through learning, with the aim of improving the recognition rate for a specific speaker,
It is possible to provide a speech recognition method that has a learning method that enables high speed, high efficiency, and automatic learning without supervision.

[Brief explanation of drawings]

Figure 1 is a diagram showing the positions of standard patterns of five vowels for two speakers, Figure 2 is a schematic diagram of phoneme set-up, and Figure 3a is a diagram showing an example of creating a standard pattern using a conventional method. Figure 3b is a diagram showing an example of standard pattern creation according to an embodiment of the present invention, Figures 4a and b are diagrams showing the relationship between phoneme learning order and misrecognition rate, and Figures 5a and b are learning FIG. 3 is a diagram showing the relationship between the number of samples and the misrecognition rate.

Claims (1)

[Claims] 1. A parameter vector of a plurality of words, syllables, and phonemes is combined for each speaker to create one parameter vector, and a set of the combined parameter vectors is clustered, and each The representative vector of the cluster is set as a set of standard patterns, and learning is performed for the speaker based on the set of standard patterns. A speech recognition method characterized by selecting a set of compatible standard patterns and identifying speech using the selected standard patterns. 2. The speech recognition method according to claim 1, wherein the standard patterns to be set are vowels.