JPH0558552B2 - - Google Patents

Abstract

PURPOSE: To restrict the existing position of a word by selecting a hypothesis that the fundamental frequency shape between supposed division points corresponds to any one of all reference shapes as a real hypothesis. CONSTITUTION: Since a pause is not generated in a level smaller than a word, the pause is detected at least and a conversation sentence is divided at the pause. Then the minimum position of fundamental frequency in the divided section is detected and the position is set up a divided candidate point. A division hypothesis supporting when each division point is a division point or when each division point is not a division point is set up and a hypothesis that the fundamented frequency shape between a divided point and a supposed point in each hypothesis corresponds to any one of all standard shapes is selected. When plural hypotheses are selected, the hypothesis having the highest reliability is selected as a real hypothesis. Thus, the division and shape of the conversation sentence is determined based upon the division point of the selected hypothesis and the obtained standard shape.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a method for understanding spoken conversational sentences (simply referred to as conversational sentences), and in particular, a method for dividing a conversational sentence in which a plurality of sentences are uttered in succession, Furthermore, the present invention relates to an apparatus for estimating the fundamental frequency shape of a spoken conversation sentence, which determines the fundamental frequency shape necessary for estimating the sentence structure of a conversation sentence.

[Prior art] Conventionally, the method for determining the fundamental frequency shape of conversational sentences has been based on a model of the fundamental frequency pattern of Japanese accent and its generation mechanism (Journal of the Acoustical Society of Japan 27-9, 445).
453, 1971) and a method of linear approximation of pitch patterns (Speech Research Group Material S47-11, 1974).

[Problems to be solved by the invention] The above-mentioned conventional technology was developed for analysis of fundamental frequency shapes in the field of speech synthesis, and if humans are involved in setting parameters etc., highly accurate estimation is possible. is possible. However, in the field of recognition and understanding, it is necessary to automatically estimate the fundamental frequency shape, but with conventional techniques, when automatic estimation is performed, accuracy deteriorates. Another disadvantage is that the amount of processing is large.

An object of the present invention is to accurately and automatically estimate the fundamental frequency shape of a spoken conversation sentence.

[Means for Solving the Problems] The fundamental frequency shows an increase at the beginning of the utterance and then gradually decreases, so it shows a triangular or "he" shape. Therefore, by finding the minimum point of the fundamental frequency, it is possible to cut out a set of fundamental frequency shapes. However, there is a problem in that a local minimum point occurs even within a set of fundamental frequencies due to speech fluctuations.

In addition, in an actual conversational sentence, the shape cut out as a unit is not only a triangle or the character ``he'', but also an interjection or a word with a short beat before the triangle. In this way, the low frequencies are connected. Furthermore, when a flat-accented phrase is connected after a raised-accented phrase and an accent combination occurs between the two, a low, flat fundamental frequency is connected after the triangle.

Therefore, the above objectives are (1) to define a standard fundamental frequency shape that takes interjections and accent combinations into consideration;
(2) Using the detected minimum point of the fundamental frequency as a dividing point, a dividing hypothesis is established assuming that each point is either a dividing point or not a dividing point. In each hypothesis, this is achieved by selecting a hypothesis whose fundamental frequency shapes between the assumed dividing points all correspond to one of the standard shapes as the true hypothesis.

[Action] Pause at a level smaller than a word
is not generated. For this purpose, first, pauses are detected and the conversation is divided at the pauses. Second, the minimum position of the fundamental frequency in the divided section is detected, and this position is set as a candidate point for division. Then, a division hypothesis is created assuming that each division point is a division point or is not a division point, and the fundamental frequency shapes between the assumed division points for each hypothesis all correspond to one of the standard shapes. Select. If multiple hypotheses are selected, the hypothesis with the highest degree of confidence is determined to be the true hypothesis. Thereby, it becomes possible to divide the conversation sentence and determine the shape based on the selected hypothetical dividing point and the obtained standard shape.

[Example] Hereinafter, an example of the present invention will be described with reference to FIGS. 1 to 3.

FIG. 1 shows a block diagram of an apparatus for estimating the fundamental frequency shape of spoken conversation sentences. Conversation audio is
After the signal is converted into an electrical signal by a microphone (not shown), the audio processing unit 1 extracts the audio power and fundamental frequency. The pause detection processing unit 2 detects a pause (for example, a silent section of 300 msec or less) from the audio power, and divides the conversation at the positions before and after the pause.
Next, the fundamental frequency minimum point detection unit 3 detects the position of the valley of the fundamental frequency, and sets this position as a division candidate point. Here, the fundamental frequency does not occur in consonant intervals such as fricatives. For this reason, the fundamental frequency extracted from conversational sentences is lacking. Therefore, the fundamental frequency minimum point detection section 3 approximates the fundamental frequency with a straight line. This approximation can be done, for example, by approximating a straight line between the fundamental frequencies separated by several points, and then finding the approximate straight line with the maximum length such that the error between the approximate straight line and the extracted fundamental frequency is less than a certain threshold. It can be achieved by asking for it. The basic frequency minimum point detecting section 3 detects the position of the valley of the approximate straight line obtained in this manner.

Next, the basic frequency division/labeling processing unit 4 first establishes a division hypothesis for each section divided by each pose. That is, a hypothesis is generated assuming that the dividing point candidate in the divided section is a dividing point and when it is not. (For example, when there is one division candidate point, there are two hypotheses: one that the entire section divided by the pose is one set of shapes, and one that consists of two sets of shapes before and after the division candidate point. (Generate a division hypothesis.) Next, using the standard fundamental frequency shape dictionary 5, the fundamental frequency shapes of the sections that are set as one shape in the hypothesis are stored in the dictionary 5. Verify whether the shape corresponds to one of the standard shapes. Then, a division hypothesis in which all the shapes of each division hypothesis correspond to one of the standard shapes is determined to be a true hypothesis. Here, if multiple division hypotheses are selected, the hypothesis with the highest degree of certainty is selected. The confidence level is, for example, an error between a group of shapes and a standard shape.

As standard shapes to be stored in the standard fundamental frequency shape dictionary 5, for example, those shown in FIG. 2 can be considered. The basic frequency of words and phrases is basically generated in the form of rising + flat + falling. Flat parts do not occur in high-headed accent clauses, so the basic shape is triangular or trapezoidal. Next, as for the decline, the undulating accent clause shows a steep decline, while the flat accent clause shows a gradual decline in the phrase component. Therefore, the basic form is further divided into two parts due to the downward slope. Furthermore, if the fundamental frequency of the first beat of a word is generated low, or if two clauses cause an accent combination, a fundamental frequency below the rising or falling slope of the basic form will be connected before and after the basic form. do. Furthermore, at the end of an interrogative sentence, a fundamental frequency that ends with a rising is generated. A straight line approximation of these shapes is shown in FIG.

The basic frequency division/labeling processing unit 4 verifies a group of shapes using the standard shape shown in FIG. . Third
The figure is an example in which the shape of each standard straight line is defined by its duration fundamental frequency fluctuation and its slope.

As described above, according to this embodiment, it is possible to automatically detect the position of the fundamental frequency shape of a group of spoken conversation sentences and determine the shape at the same time.

〔Effect of the invention〕

According to the present invention, the fundamental frequency shape can be estimated taking interjections and accent combinations into consideration, resulting in the following effects. First, in the interval estimated to be the first beat of the word connected low before the basic form of the standard shape,
If there are more syllables, all but the last syllable are always interjections, so this interjection can be deleted. In addition, in the section estimated to be the accent combination shape,
By dividing at that position, clause boundaries can be detected. Therefore, if the estimated shape is used for speech recognition, it becomes possible to limit the location of words, and performance can be improved.

Furthermore, since it is possible to automatically estimate the shape, it can also be used in accent creation in fields such as speech synthesis.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a diagram showing an example of a standard fundamental frequency shape necessary for shape estimation, and FIG. 3 shows each element of the standard shape (straight line conditions). 1 is a diagram showing an acoustic processing section, 2 is a pose detection processing section, 3 is a fundamental frequency minimum point detection section, and 4 is a fundamental frequency division/
The labeling processing unit 5 is a standard fundamental frequency shape dictionary.

Claims (1)

[Claims] 1. A fundamental frequency shape estimating device for a spoken conversation sentence that divides the spoken conversation sentence and estimates a fundamental frequency shape necessary for estimating the sentence structure of the spoken conversation sentence, comprising: A means for extracting the power and fundamental frequency, a means for detecting pauses from the extracted voice power and dividing the spoken conversation at positions before and after the pause, and using the position of the valley of the extracted fundamental frequency as a dividing point. means for detecting, means for storing in advance a standard fundamental frequency shape that takes into account at least interjections and accent combinations of the spoken conversation, and verifying the dividing lord complement point based on the stored standard fundamental frequency shape, A device for estimating a fundamental frequency shape of an audio conversational sentence, comprising: means for estimating a fundamental frequency shape of a conversational sentence;