JPH0634193B2 - Keyword extractor - Google Patents

Description

The present invention relates to an apparatus for extracting a keyword from continuously uttered speech.

[Conventional technology]

When a human hears a story, it is considered that the speech content of the other party is predicted by estimating the “topic”. Even if the speech recognition device can estimate the topic, even if it is not possible to predict the utterance content, it is possible to preselect the words by selecting the word set that was in the topic from a plurality of word sets, A recognition device with a high recognition rate can be realized. In order to estimate the topic, a word that can identify the topic in the sentence (hereinafter, such a word is referred to as a keyword) may be detected.

The technique of detecting and positioning a specific word in a sentence is called word spotting. As a word spotting method, a continuous DP method (“continuous word recognition using continuous DP” Ryuichi Oka, Voice Study Group Material S78-20) is known. FIG. 2 shows an example of the continuous DP method. In the figure, 11 is a pattern of continuously uttered input voice, 1
2 is a keyword pattern, 13 is a distance between a keyword pattern obtained as a result of continuous DP and an input voice pattern, and 14 is a matching plane. In the continuous DP method, word spotting can be performed by tracing the path on the matching plane 14 in reverse from the point of giving a distance equal to or smaller than a predetermined threshold.

[Problems to be solved by the invention]

However, the continuous DP method requires a large amount of calculation because the distance is calculated for each input frame. Therefore, if the section in which the keyword is present can be specified, the calculation amount can be reduced and the keyword can be searched efficiently.

[Means for solving problems]

The keyword extracting device of the present invention stores pitch extracting means for extracting pitch information from continuously uttered input voices, a standard pattern of keywords for estimating a topic, and a pitch change pattern indicating that the keywords are emphasized. Storing means, section extracting means for comparing the pitch information and the pitch change pattern to extract a section including the keyword, feature extracting means for converting the input voice into a series of feature vectors, and A matching unit that matches the standard pattern in a different section; and a determining unit that determines the keyword from the result of the matching.

[Action]

When a human utters a sentence, a word relating to information to be transmitted is emphasized. Therefore, it is considered that the word emphasized in the sentence is often a word that specifies the topic of the sentence, that is, a keyword. Therefore, if the highlighted portion in the sentence can be extracted, the keyword can be searched efficiently. In Japanese, it has been reported that the pitch pattern changes when a certain word in a sentence is emphasized ("Regulations of fundamental frequency pattern control in conversational sentences" Kazuhiko Iwata, Speech Study Group Material S8.
5-42). Iwata states that the change in pitch pattern depends on the emphasized word and the accent type of the preceding and following words (which syllable of the word has the accent), but basically the pitch rises in the emphasized word. . Therefore, by capturing the pitch change pattern, the emphasized portion in the sentence can be extracted.

〔Example〕

Next, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the present invention. Also,
FIG. 3, FIG. 4 and FIG. 5 are diagrams for explaining the operation of one embodiment of the present invention.

Referring to each drawing, a standard pattern of keywords and a pitch change pattern indicating that each keyword is emphasized are registered in a keyword dictionary (storage unit) 3 in advance. The feature parameter that is a standard pattern is not limited to the one that indicates the direct acoustic feature such as the output of the bandpass filter and the mel cepstrum coefficient, but may be a pattern that is symbolized by vector quantization or the like. One or more keywords should be registered for each topic. Now, it is assumed that "today" is registered with one keyword of the topic "time" and that the pitch pattern of "today" rises when stressed. FIG. 3 (a) shows the energy of the input voice "Today is a very nice day". When a voice is input, first the pitch extraction unit 1
At, the pitch is extracted. Various methods can be used as the method of pitch extraction. For example, it is a method of obtaining from the autocorrelation of the error signal of the linear prediction analysis. Fig. 3 (b)
Is the result of pitch extraction, the solid line is the case where the word "today" is not emphasized and the dotted line is the case where the word is emphasized. Therefore, the pitch is rising in “Today”. Next, the section extraction unit 2 compares the extracted pitch change pattern with the pitch change pattern of "today" registered in the keyword dictionary 3, and selects the section in which the pitch is rising, "today" as a keyword candidate. Extract as a section. Next, the characteristic extraction unit 4 obtains the characteristic parameters of the input voice. The matching unit 5 extracts the standard pattern of "today" from the keyword dictionary 3 and performs matching with the extracted keyword candidate section. Various methods can be used as a matching means. For example, the continuous D
P method. 4 and 5 show an example of matching by the continuous DP method. FIG. 4 is a distance obtained from the result of continuous DP between the keyword candidate section 21 and the standard pattern 22 shown in FIG. As a result of the matching, the decision unit 6
If a value smaller than the threshold value α is obtained, the keyword “today”
Is determined to exist. Further, when the time point when the distance takes the minimum value is t, the position of the keyword “today” can be determined by tracing the path whose end passes through the time point t on the matching plane 23.

〔The invention's effect〕

As described above, according to the present invention, a keyword can be searched for in a highlighted section in a sentence, so that the amount of calculation can be reduced and the keyword can be searched efficiently. Therefore, the continuous speech recognition apparatus can be used for estimating topics and scenes, and can perform highly accurate recognition.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a view for explaining a conventional technique, and FIGS. 3, 4, and 5 are for explaining an operation of the embodiment of the present invention. FIG. 1 ... pitch extraction unit, 2 ... section extraction unit, 3 ... keyword dictionary, 4 ... feature extraction unit, 5 ... matching unit, 6
...... Determining unit, 21 …… Keyword candidate section, 22 …… Standard pattern, 23 …… Matching plane.

Claims (1)

[Claims] 1. A memory for storing pitch extracting means for extracting pitch information from continuously uttered input voices, a standard pattern of keywords for estimating a topic, and a pitch change pattern indicating that the keywords are emphasized. Means and
Section extraction means for comparing the pitch information and the pitch change pattern to extract a section including the keyword, feature extraction means for converting the input voice into a series of feature vectors, and the standard in the extracted section. A keyword extracting device comprising: matching means for matching with a pattern; and deciding means for deciding the keyword from the result of the matching.