JPS61113099A - Voice section detecting system for voice recognition equipment - 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 speech recognition device that compares and matches a plurality of standard speech patterns and an input speech pattern, and uses the most satisfactory pattern as a recognition result. In particular, the present invention relates to a method for detecting correct speech intervals for input speech containing noise.

[Conventional technology]

In general, in speech recognition, a portion of speech input from a microphone is detected by speech section detection, and that portion is processed as a target for comparison and verification. In conventional voice section detection, voice section detection is performed by checking voice power information, etc., and if it is greater than a certain threshold value, voice exists, and if it is smaller, it is not voice.

Specific examples will be described below with reference to the drawings.

Figure 2 shows an example of a conventional speech recognition device.
1 is a microphone, 2 is an amplifier, 3 is a 16-channel filter, 4 is a power calculation unit, 5 is an interval detection unit, 6 is an input audio buffer, 7 is an audio dictionary, 8 is a matching unit, and 9 is a matching result determination unit. be.

An audio signal input from a microphone 1 is amplified by an amplifier 2 and supplied to a 16-channel filter 3. The filter 3 separates the input audio signal into predetermined bands, parameterizes each frame, and converts the signals into audio patterns. The power calculation unit 4 performs power calculation based on the audio parameters output from the filter 3.

The section detection section 5 applies a preset threshold to the power calculation result in the power calculation section 4, sets the section of the signal with the power below the threshold as a silent section, and sets the section of the signal with the power above the threshold as a sound section. As a section, only the voice pattern of the voice section is stored in the input voice buffer 6.

A large number of standard voice patterns (parameters) created in advance are registered in the voice dictionary 7, and the collation unit 8 sequentially extracts standard voice patterns from the voice dictionary 7 for each input voice pattern in the input voice buffer 6. Compare the two and calculate the distance. The matching result determination unit 9t detects the standard speech pattern with the minimum distance from the matching results and outputs it as a recognition result.

FIG. 3 shows the relationship between the noise level in the input audio signal and the threshold value, where V represents the input audio signal, E7 represents the noise level, and Sl and S2 represent the threshold levels.

The illustrated audio signal V8 shows the audio power level of the word "Funabashi", and has the strong sounds "su" and "b" in the center, followed by the weak sounds "f" and "shi".

The noise level E is in a state higher than the power level of the weak voice sounds "F'" and "S".

If the threshold level is set to Sl here, a weak speech signal can be detected, but on the other hand, there is a possibility that noise in a true silent section will be recognized as speech.

On the other hand, the threshold level is the noise level E. If S2 is set higher than S2, weak speech cannot be detected and may cause recognition failure, so it is necessary to set an appropriate threshold.

[Problem that the invention seeks to solve]

As described above, the conventional speech segment detection method has a problem with the accuracy of speech segment detection of input speech in a noisy environment, which is a major cause of lowering the recognition rate.

[Means for solving problems]

In order to solve the above-mentioned problems, the present invention first detects and determines the high-power voice parts that are not affected by noise in the input voice, and then determines the low-power voice parts that are easily affected by the noise before and after the detected voice parts. calculates the entire speech section by applying the section length collected in advance for standard speech,
The structure of the invention is to detect low-power speech parts that are susceptible to noise in a speech recognition device that compares and recognizes a plurality of standard speech patterns and input speech patterns. For standard speech, the first threshold is used to detect speech sections, and the second threshold is used to detect speech sections with strong power that are not affected by noise. The difference between the speech section detected by the threshold and the speech section detected by the second threshold is stored together with the standard speech pattern, and the speech section is detected using only the second threshold for the input speech for recognition. The first threshold value and the second threshold value stored together with the standard voice pattern are set at both ends of the detected voice section.
The feature is that the portion to which the difference section of the detection section with respect to the threshold value is added is compared and verified as the speech section of the recognition input pattern with respect to the standard speech pattern.

[Action of the invention]

As shown in FIG. 4, the present invention has a threshold 4MS + for detecting a low-level speech section and a threshold S2 for detecting a speech section reliably higher than the noise level.
A standard speech pattern is created in an environment where the noise level is lower than that of the speech section at the threshold S1, that is, the starting end t, where the speech power is weak.

and a terminal portion t8, and a central portion t where audio power is strong.

At the same time, use another threshold value S2 to find the section of the center tyo, and from these differences, calculate 1s and 1s, respectively.
．． Let's find out.

Next, in the recognition process, even if the input speech to be recognized has a high noise level as shown in FIG. 5 and the threshold SI cannot distinguish between weak parts of the speech and noise, the threshold S2 is It is possible to detect the central part t1 of a strong voice section that is not much different from the standard voice. Therefore, to the central part tffl detected in S2, we add the maximum interval in which the start end is expected to exist (the start end is called the valid interval) and the maximum interval in which the end end is expected to exist (referred to as the end valid interval). , find a longer audio storage section and store the audio in that section in a buffer.

In order to compare this longer input speech to be recognized with the standard speech, the speech section that defines the recognition range is determined as follows.

The calculations for each standard voice are 1s, 1. , is the central part t detected with the threshold value S2 for the input speech.

In addition to this, the start and end points are determined and used as a voice section.

This voice section differs depending on the standard voice.

In this way, a recognized speech section is determined for each standard speech, and recognized by comparison.

〔Example〕

The details of the present invention will be explained below with reference to Examples.

FIG. 1 is a configuration diagram of an apparatus according to an embodiment of the present invention. In the figure, 11 is a microphone, 12 is an amplifier, 13 is a 16-channel filter, 14 is a power calculation unit, 15 is an interval detection unit, 16 is a human voice buffer, 17 is a voice dictionary, 18 is a matching unit, and 19 is a matching result The determination section is shown.

The basic functions of the device of this embodiment are almost the same as those of the conventional device shown in FIG.
7's configuration and functions have been changed.

The device of this embodiment has two operating modes: a standard dictionary creation mode and a recognition processing mode.

First, in the standard dictionary creation mode, a process of creating a standard speech pattern and registering it in the speech dictionary 17 is performed. In this mode, the section detection unit 15 uses a low level SI that can detect weak speech and a level 82 that is higher than the normally assumed environmental noise level.

Place the microphone 11 in an environment where the noise level is lower than 81,
Input a predetermined standard voice.

The input audio signal is amplified by an amplifier 12, further demultiplexed by a filter 3, converted into audio parameters, and converted into an audio pattern.

The power calculation unit 14 calculates the power level of the input audio for each frame and supplies it to the section detection unit 15. The section detection unit 15 detects the power level of the input audio using two thresholds, Sl and S2, and detects the power level of the input audio using two thresholds, S1 and S2, and
The speech interval (ts +t, +t,
) and, depending on S2, the central part t,
From these, find the starting end t8 and the ending end t6, respectively.
It is registered in the speech dictionary 17 via the input speech buffer 16 along with the input speech pattern.

When all the standard voices have been registered in the speech dictionary 17, the mode is switched to recognition processing mode, and recognition processing for unknown input speech is started.

In the recognition processing mode, the threshold value of the section detection unit 15 is set to only 32. When an unknown voice is input, the section detection section 15 detects the power level of the input section using the threshold value S2, and calculates the central portion t. Furthermore, based on this t7, the start end effective section and end end effective section explained in FIG. . Note that at this time t.

is also stored as part of the data.

The matching unit 18 sequentially extracts each standard speech pattern registered in the speech dictionary 17 and matches it with the input speech pattern. At this time, the section of the input audio pattern extracted from the sofa 16 as the input audio is determined as follows. That is, the section detection unit 15 extracts the section data of the start end t5 and end section 1e from the speech dictionary 17 for each standard pattern, and combines them into the central section t□ which was previously stored together with the input speech pattern in the input speech buffer 16. Then, the voice section (1s+1.+1e) to be verified is determined. Then, from the voice pattern of the voice storage section in the input voice buffer 16, a voice pattern corresponding to this voice section is cut out and transferred to the matching section 1.

The matching section 18 sends the matching results (distance calculation results) between the input voice pattern and each standard pattern to the matching result determining section 19 for recognition determination.

〔Effect of the invention〕

According to the present invention, even if the noise level is higher than the weak power part at the beginning or end of the voice, it is possible to detect a fairly accurate @Ashi section, and it is possible to reduce misrecognition caused by the influence of noise. can.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of a device according to an embodiment of the present invention, Fig. 2 is a block diagram of a conventional device, Fig. 3 is an explanatory diagram of the relationship between noise level and threshold, and Fig. 4 is a section when creating a speech dictionary. FIG. 5 is an explanatory diagram of the detection process. FIG. 5 is an explanatory diagram of the section amount detection process during the recognition process. In the figure, 11 is a microphone, 12 is an amplifier, 13 is a filter,
14 is a power calculation section, 15 is a section detection section, 16 is an input voice buffer, 17 is a speech dictionary, 1 is a collation section, and 19 is a collation result determination section. Patent applicant Fujijff! Representative Patent Attorney Co., Ltd. Fumihiro Hase (1 other person) (G) U
J U Noriguchi Q table

Claims (1)

[Claims] In a speech recognition device that compares and matches multiple standard speech patterns and input speech patterns to recognize them, there is a first threshold that can detect low-power speech parts that are susceptible to noise, and a first threshold that can detect low-power speech parts that are susceptible to noise. A second threshold that can detect strong speech parts is provided, and for standard speech, the first threshold is used to detect speech sections, and the speech sections detected using the first threshold and the second threshold are detected. The difference between the detected speech section and the standard speech pattern is stored together with the standard speech pattern, and the speech section is detected using only the second threshold for the input speech for recognition, and the standard speech pattern is added to both ends of the detected speech section. Speech section detection characterized by comparing and comparing a portion to which a difference section of a detection section based on a first threshold value and a second threshold value, which are stored together, is added, as a speech section of a recognition input pattern with respect to the standard speech pattern. method.