JPS63173100A - Keyword extractor - 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 [Field of Industrial Application] The present invention relates to a device for extracting keywords from continuously uttered speech.

[Conventional technology]

When Taiko listens to a conversation, it is thought to predict the content of the other person's utterances by estimating the ``topic''. If the speech recognition device can also estimate the topic, even if it is not possible to predict the content of the utterance, it will be possible to make a preliminary selection of words by selecting a word set that matches the topic from among multiple word sets. A recognition device with a high recognition rate can be realized.

In order to estimate the topic, it is sufficient to detect words (hereinafter such words will be referred to as keywords) that can specify the topic in the sentence.

The technology for detecting and positioning specific words in a sentence is called word spotting. The method for word spotting is the continuous DP method ("Continuous word recognition using continuous DP" Takashi Oka-1 Speech Research Group Material 578-20)
It has been known. The second factor shows an example of the continuous DP method. In the figure, 11 is a pattern of continuously uttered input speech,
12 is the keyword pattern, 13 is the distance between the keyword pattern obtained as a result of continuous DP and the input speech pattern, and 14 is the matching plane. Word spotting can be done by retracing the path above step 14.

[Problem that the invention seeks to solve]

However, in the continuous DP method, distance calculation is performed for each frame of input, which requires a large amount of calculation. Therefore, if the section in which a keyword is located in a sentence can be identified, the amount of calculation can be reduced and keywords can be searched efficiently. .

[Means for solving problems]

The keyword extraction device of the present invention stores pitch extraction means for extracting pitch information from continuously uttered input speech, a standard pattern of keywords for estimating a topic, and a pitch change pattern indicating that the keyword is emphasized. a storage means for comparing the pitch information and the pitch change pattern to extract a section including the keyword; a feature extraction means for converting the input speech into a sequence of feature vectors; and a determining means that determines the keyword from the result of the matching.

[Effect]

When humans utter sentences, words related to the information they want to convey are emphasized. Therefore, it is considered that the words emphasized in a sentence are often words that specify the topic of the sentence, that is, keywords. Therefore, if you can extract the emphasized parts of a sentence, you can search for keywords efficiently. In Japanese, it has been reported that the pitch pattern changes when a certain word in a sentence is emphasized ("Conversation Sentences"). Regarding the control rules for fundamental frequency patterns in ``Iwa 1) Aiseki, Speech Study Group Material 58
5-42), Iwata states that changes in pitch patterns depend on the word being emphasized and the accent type of the words before and after (which syllable of the word is accented), but basically Pitch rises in words. Therefore, by capturing the pitch change pattern, it is possible to extract the emphasized part of the sentence.

〔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.

Further, FIGS. 3, 4, and 5 are diagrams for explaining the operation of an embodiment of the present invention.

Referring to each figure, a standard pattern of keywords and a pitch change pattern indicating that each keyword is emphasized are registered in advance in the keyword dictionary (storage unit) 3, and the characteristic parameters of the standard pattern are filtered by a bandpass filter. The keywords are not limited to those indicating direct acoustic features such as the output of , mel-cepstral coefficients, etc., but may also be patterns symbolized by vector quantization, etc. One or more keywords shall be registered for each topic. It is now assumed that "Today" is registered as one of the keywords of the topic "Season", and that it is remembered that the pitch pattern of "Today" rises when it is emphasized. Figure 3 (a
> indicates the energy of the input voice "It's very nice weather today." When the voice is input, the pitch is first extracted in the pitch extraction unit 1. Various methods can be used for pitch extraction. For example, there is a method of obtaining the pitch from autocorrelation of an error signal of linear predictive analysis. FIG. 3(b) shows the results of pitch extraction, where the solid line shows the case when the word "Today" is uttered without emphasis, and the dotted line shows the case when it is uttered with emphasis. Therefore, the interval extraction unit 2 compares the extracted pitch change pattern with the pitch change pattern of "Today" registered in the keyword dictionary 3, and detects the rise in pitch. 0 to extract “Today” as a keyword candidate interval.
Next, the feature extraction section 4 obtains the above-mentioned feature 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, such as the continuous DP method described above.

FIGS. 4 and 5 are examples of matching using the continuous DP method. Figure 4 shows the keyword candidate section 21 shown in Figure 5.
This is the distance obtained from the result of continuous DP between the standard pattern 22 and the standard pattern 22. The determining unit 6 determines that the keyword "today" exists if a value smaller than the threshold α is obtained as a result of matching. Further, assuming that the time point at which the distance takes the minimum value is t, the position of the keyword "today" can be determined by tracing a path on the matching plane 23 whose end passes through time t.

〔Effect of the invention〕

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

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a diagram for explaining the prior art, and FIGS. 3, 4, and 5 are for explaining the operation of an embodiment of the present invention. This is a diagram. 1... Pitch extractor, 2... Section extractor, 3...
Keyword dictionary, 4...Feature extraction unit, 5...Matching unit, 6...Determination unit, 21...Keyword candidate section, 22...Standard pattern, 23...Matching plane. 1st

Claims (1)

[Claims] pitch extraction means for extracting pitch information from continuously uttered input speech; storage means for storing a standard pattern of keywords for estimating a topic; and a pitch change pattern indicating that the keyword has been emphasized; a section extraction means for comparing information with the pitch change pattern and extracting a section including the keyword; a feature extraction means for converting the input voice into a series of feature vectors; What is claimed is: 1. A keyword extraction device comprising: a matching means for performing matching; and a determining means for determining the keyword from the result of the matching.