JPH03160499A - Speech recognizing device - Google Patents

Abstract

PURPOSE:To avert the erroneous recognition by noises without executing sound volume control of an acoustic apparatus which is a noise source by prohibiting the comparison processing of an identifying section or nullifying the output thereof when a voice pattern and an acoustic pattern coincide. CONSTITUTION:A sound analyzing section 5 which analyzes ambient sounds is provided separately from a voice analyzing section 1 for voice recognition and a means 7 for deciding whether the sound inputted to the voice analyzing section 1 is the ambient sound or not according to the degree of coincidence of the voice for voice recognition and the sound outputted from the acoustic apparatus is provided. If the voice analyzing section 1 is regarded to analyze not the voice but the ambient noise by the deciding means 7, the voice recognition to the voice pattern of this time is not executed to avert the erroneous recognition of the voice. The inconvenience to recognize the ambient noise is averted in this way and the erroneous recognition is lessened.

Description

[Detailed description of the invention] (a) Industrial application field The present invention enables effective speech recognition even in the vicinity of ambient noise sources such as audio equipment that outputs sounds such as voice and music, such as TVs and stereos. The present invention relates to a speech recognition device.

(B) Prior Art Conventional speech recognition devices have the risk of causing frequent recognition errors when placed near audio equipment such as a stereo, as the output sound of this equipment becomes ambient noise. Particularly, when such equipment itself is operated by a voice recognition device, there is a risk that the voice recognition device may operate on its own due to the voice or music emitted by the audio device.

Therefore, in order to prevent such malfunctions, a speech recognition device has been proposed in which, for example, when inputting a voice, the output sound of the audio device is reduced during the voice input period [JP-A-63 -29755 publication]. However, such a device not only has the inconvenience of having to interrupt the audio and music from the audio equipment, but also has the following problems.

That is, the above-mentioned device requires a switch operation to specify the audio input period, and this switch operation is avoided and the audio input period is detected based on the sound pressure level of the audio input to the microphone. However, it is still not possible to completely avoid the possibility that the sound from the audio equipment may interfere with this detection operation.

(c) Problems to be Solved by the Invention The present invention provides a speech recognition device that can avoid erroneous recognition due to noise without controlling the volume of an audio device that is a noise source.

(2) Means for Solving the Problems The speech recognition device of the present invention includes a speech analysis section that separates input speech and a first pattern creation section that produces speech patterns from the speech analyzed by the speech analysis section. , an identification unit that compares the voice pattern created in the first pattern creation section with a standard voice pattern prepared in advance and outputs a recognition result; an acoustic analysis section that separately analyzes ambient sound; a second pattern creation section that creates an acoustic pattern from the ambient sound analyzed by the acoustic analysis section; the acoustic pattern created by the second pattern creation section and the first pattern; A determining unit is provided to determine whether the audio pattern created by the creating unit matches, and the determining unit determines whether the audio pattern created by the first pattern creating unit matches the acoustic pattern created by the second pattern creating unit. If it is determined, the comparison process of the identification section is prohibited or its output is invalidated.

(E) Function According to the speech recognition device of the present invention, an acoustic analysis section for analyzing ambient acoustics is provided separately from the speech analysis section for speech recognition,
A means is provided for determining whether the sound input to the voice analysis section is ambient sound based on the degree of coincidence between the voice for voice recognition and the sound output from the audio equipment, and this determination means determines whether the sound input to the voice analysis section is ambient sound or not. When it can be assumed that the analysis is not a voice but ambient noise, voice recognition is not performed for the voice pattern at this time, so that erroneous voice recognition can be avoided.

(F) Embodiment FIG. 1 shows a block diagram of the structure of the speech recognition device of the present invention. In the figure, reference numeral 1 denotes a voice analysis unit, which cuts out the voice time domain from the acoustic signal input by the microphone 9 set toward the speaker, performs frequency analysis on the voice signal in this domain, and performs spectrum analysis. Or perform voice analysis such as ceps drum analysis. Reference numeral 2 denotes a first pattern creation section, which samples and normalizes the analysis results from the speech analysis section 1 to obtain a speech pattern expressed as a vector pattern, a cepstrum pattern, or the like. 3 is a voice pattern storage section;
A plurality of speech patterns extracted by analyzing standard specific or unspecified speech in advance are stored as standard patterns for recognition. 4 is an identification unit that sequentially matches the audio pattern created by the first pattern creation unit 2 with a plurality of standard patterns stored in advance in the audio pattern storage unit 3, and recognizes the most similar one. This is where the results will be output.

The structure of the groove or the conventional device is the same as described above, and the feature of the device of the present invention is that the following structure is added.

Reference numeral 5 in the figure denotes an acoustic analysis section, which analyzes the audio or music output from the audio equipment 8 such as a car stereo, and uses the same spectral analysis or cepstral analysis as the above-mentioned audio analysis section 5. be done.

Reference numeral 6 denotes a second pattern creation section, which obtains an acoustic pattern from the sound analyzed by the acoustic analysis section 5 through sampling normalization similar to that of the first pattern generation section 2 described above.

Reference numeral 7 denotes a voice determination section, which compares both patterns created simultaneously by the first pattern creation section 2 and the second pattern creation section 6 to determine whether or not to perform the identification process of the identification section 4.

The operation of the apparatus of the present invention will be explained below, focusing on the processing of the acoustic analysis section 5, second pattern creation section 6, and voice determination section 7.

First, the speech analysis section 1 and the acoustic analysis section 5 will be explained.

These voice analysis section 1 and acoustic analysis section 5 have the same structure, and simultaneously sample at fixed time intervals (eg, 5 msec). The structure of these analysis sections 1.5 is similar to the analysis section of the conventional device, as shown in FIG.
Bi-emphasis 102, bandpass filter 103,
It is composed of an AD converter lO4, a voice detection circuit 105, a buffer 106, and the like. That is, in the voice analysis section 1, the sound input from the microphone 9 is amplified by the amplifier 101,
A pre-emphasis 102 applies Takagi emphasis, a bandpass filter 103 performs filter analysis, and an AD converter 104 converts the signal into a digital value. On the other hand, the voice detection circuit 105 checks the beginning of the voice from the data converted into digital values by the AD converter 104, and if it exceeds the cutout threshold, the data is buffered by 1.
Save in 06. At the same time, the start of audio processing is transmitted to the acoustic analysis section 5.

Basically, when the voice analysis section 1 and the same sound analysis section 5 receive the notification of the start of voice work from the voice analysis section 1, they receive the acoustic signal output from the acoustic output device 10 and receive the acoustic signal. The analysis data analyzed in the same manner as described above is loaded into the buffer 106.

Next, while continuing the analysis, the speech analysis unit 1 detects the ending of a word in the speech detection circuit 105, and captures the data into the buffer 106 until it is determined that it is the ending of a word.When it is determined that it is the ending of a word, the speech analysis unit 1 ends the capturing and detects the sound. The end of the import is transmitted to the analysis unit 5. In this way, the acoustic analysis section 5 receives the notification of the end of the capture from the voice analysis section 1 and finishes analyzing the sound from the car stereo, and finishes capturing the buffer 106.

As described above, the analysis times of the voice analysis section 1 and the acoustic analysis section 5 can be synchronized.

Although both of these analysis sections are shown as separate 7) one-domain structures, the digital circuit portions can be used in common in a time-division manner.

Next, the first pattern creation section 2 and the second pattern creation section 6 will be explained.

These pattern making sections 2 and 6 can also adopt the pattern making bottom section of the follow-up device, and both can be realized with the same structure. In the first pattern creation section 2, the analysis result analyzed by the speech analysis section 1 is converted into a speech pattern suitable for speech recognition.

On the other hand, the analysis result analyzed by the acoustic analysis section 5 is converted into an acoustic pattern by the second pattern generation section 6 using the same method as the first pattern generation section 2. Therefore, these pattern creation units may also use the same hardware in a time-sharing manner b).

Furthermore, the voice determination section 7 will be explained. This audio determination section 7 performs matching between the audio pattern created by the first pattern creation section 2 and the acoustic pattern created by the second pattern creation section 6. This ma-notching usually performs linear matching to calculate the degree of similarity between two notchings. Based on this degree of similarity and a predetermined threshold, it is determined whether or not it is a voice.

That is, when the degree of similarity is smaller than the threshold value and the two are similar, this is determined to be a match, and the audio pattern can be considered to be caused by the sound from the car stereo, so in this case, the processing operation of the identification unit 4 is prohibited.

Alternatively, in this case, the output of the identification result from the identification unit 4 may be blocked and invalidated. On the other hand, if the degree of similarity is greater than the threshold and the two are not similar, it is determined that there is a mismatch, and it can be assumed that the speaker has input voice into the microphone. The chest signal is transmitted to the identification section 4, and the identification section 4 performs an identification operation. Therefore, no matter how high the sound output from the sounder lip 8 is and the sound wraps around to the microphone 9 by disabling the operation or disabling the output of the identification unit 4, the sound will not be input until another microphone 9 is input. , voice recognition will not be possible.

First, only after receiving the judgment that the result of the judgment section 7 represents a voice input, the identification section 4 sequentially extracts patterns from the voice pattern storage section 3 and performs matching, and selects the most similar pattern. Output as recognition result.

Further, a specific example will be given and an explanation will be added.

As a prerequisite for this embodiment, it is assumed that the object of operation control by voice recognition is a car stereo, which is an audio device 8 that emits ambient noise. In this case, a microphone installed near the car stereo's speakers may be used to transmit the acoustic signal from the car stereo to the acoustic analysis section 5, but a wire is connected directly from the audio output terminal of the car stereo to the input of the acoustic analysis section 5. If you do that, you won't need the microphone mentioned above.

First, a case will be described in which the audio analysis section 1 cuts out audio based on music from the car stereo (8).

Here, if the amplifier of the audio analysis section 1 is set to a level that prevents the audio analysis section 1 from extracting the sound pressure of normal music in the audio time domain, it will be possible to analyze the music. This will reduce the possibility of it being lost. In this way, when a loud sound is output from the car stereo (8), this is input to the microphone 9 set for the driver, and the audio analysis section l starts capturing the sound from the car stereo (8). At the same time, the acoustic analysis section 5 also starts capturing.

Then, the analysis continues and the ending of the word is checked, and when the ending of the word is detected, the speech is judged by the speech determining section 7. In this case, since the input to the audio analysis section 1 is only from the car stereo 8, the audio pattern created by the first pattern creation section 2 and the second
Since the acoustic patterns created by the pattern creation unit 6 match, the degree of similarity increases, and the voice determination unit 7 determines that the sound is not the driver's voice for voice recognition. Here, the output of the car stereo (8) does not match perfectly due to a time error until it is input to the microphone 9, but this error can be reduced by adding a time rrI delay circuit to the acoustic analysis section 5. Is possible.

Furthermore, it is also possible to absorb errors by using linear shift matching or the like for matching in the voice determining section 7.

Next, a case will be described in which the car stereo (8) is controlled by voice while outputting sound.

The voice pattern storage section 3 contains "Eject", "Bray",
It is assumed that patterns such as "stop", "rewind r", and "fast forward j" are stored that have been analyzed in advance for voice control of a car stereo.When the driver speaks "stop" into the microphone 9, the voice analysis section 1 For example, a mixture of music from the car stereo (8) and the driver's voice "stop" from the microphone 9 is analyzed. Voice detection circuit l0 of voice analysis unit l
5 detects the beginning of the word "stop" and starts capturing audio, and simultaneously transmits the start of capturing to the acoustic analysis unit 5. Upon receiving this, the acoustic analysis section 5 analyzes the output sound of the car stereo (8) until the speech detection circuit 105 of the speech analysis section 1 detects the ending of the word, and captures this.

When the analysis in the voice analysis section 1 and the acoustic analysis section 5 is completed in this way, the first pattern generation section 2 and the second pattern generation section 6
Each is converted into a pattern.

The converted pattern is determined by the voice determination unit 7 to determine whether it is voice or not, but since the pattern analyzed by the voice analysis unit 1 is a mixture of the driver's voice and the sound of the car stereo (8), It did not match the pattern analyzed by the acoustic analysis unit 5, and it was determined that there was a voice input from the driver.
A signal is transmitted to the identification unit 4 to start the identification operation. The identification unit 4 first imports the "eject" pattern from the audio pattern storage unit 3 and performs matching, then matches "play" and "stop" sequentially until the end of the pattern, and selects "stop" with the highest degree of similarity. is output as the recognition result. As a result, the operation of the car stereo (8) can be stopped as intended by the driver.

Here, the pattern created in the first pattern creation section 2 was used as is, but the identification section subtracted the pattern created in the second pattern creation section from the pattern created in the first pattern creation section 2. When used in the processing in step 4, a pattern consisting only of the voice base part with the noise part canceled out can be obtained, and thereby higher recognition performance can be obtained.

In the above description, the voice control target of the voice recognition device itself is a source of ambient acoustic noise, but the present invention is not limited to this. For example, the voice control target of the voice recognition device is an ambient acoustic noise source. The present invention can be applied even when the opening/closing operation is voice-controlled.

(G) Effects of the Invention According to the speech recognition device of the present invention, as long as there is no voice input from the speaker, it is possible to avoid the inconvenience of recognizing ambient noise, and to reduce misrecognition.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the structure of the speech recognition apparatus of the present invention, and FIG. 2 is a block diagram showing the structure of the main part of the apparatus of the present invention shown in FIG. 1... Voice analysis section, 2... First pattern creation section, 3.
. . . Audio pattern storage unit, 4 . . Identification unit, 5 . ·microphone.

Claims (3)

[Claims] (1) A voice analysis section that analyzes input voice; a first pattern generation section that creates voice patterns from the voice analyzed by the voice analysis section; and a voice pattern created by the first pattern generation section that is prepared in advance. In a speech recognition device that has an identification section that outputs a recognition result by comparing it with a standard speech pattern that has been analyzed, an acoustic analysis section that analyzes ambient sound separately from the speech analysis section, and an acoustic analysis section that analyzes ambient sound a second pattern creation section that creates an acoustic pattern from ambient sound; and a determination section that determines whether the acoustic pattern created by the second pattern creation section matches the audio pattern created by the first pattern creation section. If the determination unit determines that the audio pattern created by the first pattern creation unit and the acoustic pattern created by the second pattern creation unit match, the comparison process of the identification unit is prohibited or the output thereof is A voice recognition device characterized by disabling the. (2) The speech recognition device according to claim 1, wherein the analysis time domain of the acoustic analysis section is synchronized with the analysis time domain of the speech analysis section. (3) A speech recognition device characterized in that the acoustic analysis section is directly input with an output signal of an audio device that affects the speech input to the first speech analysis section.