JPS59121099A - Voice section detector - 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 is directed to a voice that can reliably detect the voice section of an input voice signal that is continuously uttered without being affected by so-called tongue clicks or breathlessness. The present invention relates to a section detection device.

[Technical background of the invention and its problems]

When recognizing continuously inputted speech, it is very important to accurately detect speech sections. Conventionally, this voice section detection has been performed roughly in the following manner. That is, the first method focuses on the power of the input signal and uses two threshold values 1-1. One threshold value is used to reliably detect a section where voice is present.1. Then, the other threshold is used to find an accurate speech interval.

However, with this method, there is a problem in that even high-power pulsed noise may be erroneously determined as speech.

On the other hand, the second method utilizes the fact that the number of zero crossings of the input signal and its spectral outline are different from background noise, and uses characteristic parameters that can be distinguished from noise to determine speech intervals. There is. However, this method has not been found to have parameters that can definitively characterize various voices, including voiced and unvoiced sounds, so it has been improved slightly compared to the method @1 described above. It's just recognized.

Therefore, recently, consideration has been given to performing phoneme classification using frequency characteristic parameters of input speech and performing speech segment detection using the power information of the input signal in combination. In other words, the frequency structure of phonemes is registered in advance as knowledge, and the fact that the frequency structure of noise is not similar to that of phonemes is utilized. Classification is performed, and based on the time series of phoneme labels obtained through this classification, it is determined whether or not it is a speech section.

However, since this phoneme classification only examines the frequency structure of the time in acoustic analysis, for example, the so-called lisp sounds that have a frequency structure similar to the phonemes /l f/ and /B/, and the phonemes /h/ There were problems such as misjudgment of breath noise with a similar frequency structure and incorrect phonological labeling. If an error is introduced in the phoneme labeling in this way, a problem arises in that the noise cannot be removed even if the temporal structure of the phoneme label is examined in subsequent processing. As described above, in the past, it was extremely difficult to correctly determine and detect the voice section of a continuously input voice signal.

[Purpose of the invention]

The present invention takes these circumstances into consideration: 1. It was made in
The purpose of this is to correct the speech interval in the input signal without being influenced by noise such as tongue clicks or breathlessness. <
The object of the present invention is to provide a voice section detection device that can detect a voice section stably.

[Summary of the invention]

The present invention registers in advance the time structure of feature parameters in subintervals including the start and end points of a speech section as a standard pattern, and combines the time series of feature parameters obtained by analyzing an input signal with the L standard. The degree of similarity with each pattern is calculated, and the start point of the voice section in the input signal is determined and detected according to the degree of similarity.

〔Effect of the invention〕

Thus, according to the present invention, not only the frequency feature structure of the start and end points of a speech interval but also the time structure is given as a standard pattern as a feature parameter level, so that the frequency structure is similar to that of speech. However, it is possible to accurately and stably determine the start and end points of a speech interval by excluding noises such as tongue clicks and breath leaks that have different temporal structures with high precision.1. This makes it possible to detect voice sections in the input signal.

Therefore, great practical effects can be achieved, such as making it possible to reliably perform speech recognition according to correctly detected speech sections.

[Embodiments of the invention]

Hereinafter, one embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a schematic configuration diagram of the embodiment apparatus, and I is an acoustic analysis section. This acoustic analysis section 1 analyzes the audio range from 2 to 1, for example.
It consists of a well-known filter bank consisting of a plurality of band-pass filters divided into about 8 filters, and spectrally analyzes the input signal to obtain characteristic parameters corresponding to the frequency at each time and hold them. Through this acoustic analysis section 1
．． The feature parameter time series of the input signal determined by the input signal is given to the similarity calculation unit 2, and the similarity with a standard pattern registered in advance in the standard pattern storage unit 3 is calculated. Therefore, the standard pattern registered in advance in the first storage unit 3 is given as a time series of characteristic parameters of a partial section including the start time and end time of the speech section of the standard voice. In addition, in the temporal sampling method of the characteristic parameters forming this standard pattern, the characteristic parameters may be extracted continuously for a predetermined analysis time, but they may also be determined discontinuously; may be positioned at the center of the partial interval in the time direction.

However, if we focus on the starting point, in the case of Japanese speech, approximately 100 syllables have different speech patterns.
When frequency features are extracted by band division, there is no need to particularly distinguish between front vowels and back vowels. Therefore, in this case, by preparing about several dozen types of standard patterns, it is possible to detect the starting point of the Japanese input speech with sufficient reliability. In addition, if the idioms that the device handles are predetermined, you can simply prepare standard patterns for each idiom.

That purpose is achieved. On the other hand, regarding the standard pattern in the subinterval that includes the terminal point, it is important to note that the terminal side of the voice is a vowel due to its structure, and in the case of devoicing, it is necessary to distinguish 1. Since there are some vowels that are not included, it is only necessary to prepare a few standard patterns.

The similarity calculation unit 3 calculates the similarity between each standard pattern thus determined and the characteristic parameter time series of the input signal at a given time. This similarity calculation may be performed using, for example, a composite similarity calculation method.
It is done in a pattern-based manner. Therefore, the similarity calculation unit 3 performs this similarity calculation for each of the plurality of standard pacusos at each analysis time of the input signal, and thereby the time series of the similarity for each standard pattern is obtained. ing. The determining unit 4 inputs the time series of such similarities and determines the point in time at which a similarity value exceeding a predetermined threshold for each standard pattern and having a maximum value is obtained as the end point of the speech section, that is, the starting point. 17 points or terminal points are detected. It goes without saying that the starting point or the ending point is determined independently from each other by the standard pattern of the partial section including the starting point and the standard pattern of the partial section including the ending point. In this way, the start and end points of the voice are detected, and these indicate the voice section in the input signal. Note that if there are a plurality of start points and end points, a voice section is indicated by the combination thereof.

FIG. 2 is a control flow showing the above-mentioned series of interval detection processing, and shows the period during which an audio signal is being generated. Characteristic parameters of the input signal are sequentially extracted at each predetermined analysis time (step 11). Then, the time series of the above feature parameters is extracted at each time interval (Step 12), and the degree of similarity between the time series and the standard pattern is calculated (Step 1).
3). Thereafter, the starting point and ending point of the voice section may be determined in accordance with this similarity value (step 14).

In other words, as shown in FIG. 3, the time series of the input signal is 1.
For each partial section, whether it is a starting point, a terminal point, or something else can be determined and detected by matching with a standard pattern.

Note that the present invention is not limited to the above embodiments. For example, when using a standard pattern in which the end points are shifted from the temporal center of the partial section, it goes without saying that the position of the end point must be determined from the partial section in which the melt point is detected by correcting the above deviation. Also, from the power of the input signal, a portion that is clearly a silent section is detected, a portion excluding the detected silent section is extracted as a speech candidate section, and the similarity calculation described above is performed only for this section. , calculation l
can be reduced. In addition, the present invention can be implemented with various modifications without departing from the gist thereof.

As explained above, according to the present invention, the time structure of the feature parameters of the subinterval including the end points is provided as a standard pattern, pattern matching is performed with the time series of the feature parameters of the input signal, and the similarity is calculated based on the similarity. Since it detects the starting point of the speech section, it is possible to correctly detect the speech section from its temporal structure without being affected by noise such as tongue clicks or breath leaks, and the detection accuracy is high and stable. be. Therefore, it becomes possible to correctly determine the existence section of continuously uttered and inputted speech, and to perform the recognition processing effectively, resulting in a great practical effect.

[Brief explanation of drawings]

FIG. 1 is a schematic configuration diagram of an apparatus according to an embodiment of the present invention, FIG. 2 is a diagram showing the flow of voice section detection processing according to the embodiment, and FIG. 3 is a diagram showing the relationship between an input signal and a partial section. be. I...Acoustic analysis unit 2...Similarity calculation unit 3...Feature parameter storage unit 4...Judgment unit 1 agent Patent attorney Takehiko Suzue 1st case 3 Figure-70 (

Claims (3)

[Claims] (1) An acoustic analysis unit that analyzes an input signal at regular intervals to obtain its characteristic parameters, and a storage unit that stores the time structure of characteristic parameters in a partial interval including the start and end points of a voice interval as a standard pattern. , means for calculating the degree of similarity between the standard pattern and the characteristic parameter time series of the input signal, and means for determining and detecting the start and end points of the speech section of the input signal from the calculated similarity values. A voice section detection device characterized by comprising: (2) The speech section detection device according to claim 1, wherein the similarity calculation is performed on a time series of feature parameters of the speech candidate section determined from the power of the input signal. (3) The time structure of the feature parameters is based on the frequency/time structure of the voice in the partial interval including the start and end points of the voice. 2. A voice section detection device according to claim 1, wherein the voice section detection device represents a voice pattern.