JP2000163098A - Voice recognition device - Google Patents

Abstract

(57) [Summary] [PROBLEMS] The conventional speech recognition device has a problem that a speech section cannot be correctly detected in a high-noise environment, and the recognition rate decreases. In addition, the operation of the trigger button and the like is required at the time of voice input, and the operation is troublesome. SOLUTION: A mixed noise component is removed using a noise spectrum estimated from a non-voice section, and a voice section is detected by using power of only a voice band as voice power.
Furthermore, a final recognition result is output using the likelihood difference between the highest word of the likelihood calculation result and the next candidate word.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a high noise and large swing environment such as a vehicle mainly used on rough terrain, such as a vehicle used in a construction site or a disaster area or a defense vehicle. The present invention relates to a voice recognition device used to operate an electronic device.

[0002]

2. Description of the Related Art Vehicles mainly used on rough terrain, such as vehicles used in construction sites and disaster areas, and defense vehicles, generally have a large noise such as engine noise and have a quieter interior than ordinary passenger vehicles. Inferior. In addition, the use of irregular terrain such as a disaster area or a training ground is accompanied by a large swing, and in many cases, both hands are not free to support the body. In such a situation, it may be difficult to operate a device having buttons, switches, and the like. 2. Description of the Related Art In recent years, the spread of car navigation systems has been remarkable due to the development of car electronics technology, and even in such vehicles, it is desired to mount a system such as a car navigation system or an information display device. However, as described above, in a severe use environment such as a disaster area dispatch vehicle or a defense vehicle, it is difficult to operate complicated buttons and switches by hand. Therefore, it is convenient if various electronic devices can be operated by voice without using hands. Therefore, there is a demand for a voice recognition device that can be used in a vehicle used in a harsh environment with high noise and large swing. FIG. 10 shows an installation state of a voice recognition device installed in a general vehicle, 30 denotes a voice recognition device, 40 denotes a controller such as a car navigation operated by the voice recognition device, and 50 denotes a monitor such as a car navigation. Here, 30 speech recognition devices are shown in detail in FIG. In FIG. 9, 1a is a microphone for inputting sound, 1b is a microphone for picking up environmental noise, 11 is a trigger button operated when inputting sound, 2 is acoustic analysis means for calculating input power or the like from an input signal, 4 is Noise spectrum estimating means for estimating a noise spectrum from noise data, 5 means for removing noise components from an input signal using the estimated noise spectrum, and 3 means detecting a voice section using input power from which noise components have been removed. 6 is a word feature vector extraction means for extracting a feature vector from the removed speech data, 7 is a registered word dictionary storing feature vectors of registered words to be recognized, and 8 is a register of extracted feature vectors. This is a speech recognition unit that compares the likelihood with a word, calculates its likelihood, and outputs the highest word as a recognition result.

Conventionally, in voice recognition under high noise, a voice is input by pressing a trigger button 11 when voice is input in order to distinguish between environmental noise and input voice. At this time, voice is acquired by the microphone 1a and noise is acquired by the microphone 1b, and the input power, power spectrum, and the like are calculated by the acoustic analysis unit 2. The noise spectrum estimating means 4 estimates the noise spectrum by predicting the power for each frequency component of the power spectrum. The noise removing unit 5 uses the estimated noise spectrum to remove a noise component from the power spectrum of the input signal acquired by the microphone 1a to calculate a voice spectrum. Also, the power is calculated by integrating the power from the voice spectrum. The voice section is detected by the voice section detection means 3 using the calculated voice power. A word feature vector is extracted from the speech section data detected by the word feature vector extraction means 6. The registered word dictionary 7 stores the feature vectors of the registered words. When the speech section is determined, the likelihood between the word feature vector extracted by the recognition processing unit 8 and the feature vector of the registered word is calculated. The top word of the calculated likelihood is output as a recognition result.

[0004]

In a conventional speech recognition apparatus, a microphone for acquiring only environmental noise and two microphones for simultaneously acquiring speech and environmental noise are required to remove noise components. If both microphones cannot acquire in-phase environmental noise, noise will be added to the audio signal, which will cause a problem such as a significant decrease in the recognition rate. Was. In addition, under noise, it is necessary to input a sound by operating a trigger button or the like in order to distinguish the sound from the noise, which has a disadvantage that the operation is troublesome.

The present invention has been made to solve such a problem, and an object of the present invention is to provide a voice recognition device capable of performing stable recognition even under noise.

[0006]

According to a first aspect of the present invention, there is provided a voice recognition apparatus comprising: a microphone for inputting voice; an acoustic analysis unit for calculating a power spectrum from an input signal and outputting a total power of a voice band; Voice section detecting means for detecting a voice section by using a noise section, a noise spectrum estimating means for estimating a noise spectrum from data of a non-voice section, a noise removing means for removing a noise component from data of a voice section, and voice data removed. Word feature vector extraction means for extracting a feature vector from a registered word dictionary storing feature vectors of registered words to be recognized, and comparing the extracted feature vector with the registered words, calculating the likelihood thereof, and Voice recognition means for outputting as a recognition result when the likelihood difference with the candidate is equal to or greater than a set threshold value.

[0007] A speech recognition apparatus according to a second aspect of the present invention comprises:
A microphone for inputting voice, acoustic analysis means for calculating a power spectrum from an input signal and outputting the total power of a voice band, voice section detecting means for detecting a voice section using voice band power, and data for a non-voice section. A noise spectrum estimating means for estimating a noise spectrum from the speech data, a noise removing means for removing a noise component from data of a voice section, a word feature vector extracting means for extracting a feature vector from the removed voice data, and a registered word to be recognized. A registered word dictionary that stores feature vectors, a speech recognition unit that compares the extracted feature vectors with the registered word dictionary, calculates the likelihood thereof, and outputs the top few words as a recognition result, and a sequence of the top few words output And a keyword determining means for determining a keyword of the voice switch from.

[0008] Further, a speech recognition apparatus according to a third aspect of the present invention comprises:
In the speech recognition apparatus according to the first or second aspect, an SNR determination unit that controls a recognition process based on an average power of a voice section frame and an average power of a noise frame detected by the voice section detection unit according to an SN ratio is provided. Things.

[0009] A speech recognition apparatus according to a fourth aspect of the present invention comprises:
As a power used for detecting a voice section in the voice recognition device according to the first or second aspect of the present invention, a main frequency is calculated from an estimated noise spectrum, and a power spectrum of input data excluding a vicinity of a main frequency component of noise is excluded. Means for detecting a voice section using the sum of the powers.

[0010] A speech recognition apparatus according to a fifth aspect of the present invention includes:
The voice recognition device according to the first or second aspect of the present invention includes means for detecting a voice section by using a sum of powers of voice bands in a power spectrum obtained from a difference between a power spectrum of input data and a power spectrum of noise estimated. It is provided.

[0011]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiment 1 FIG. 1 shows a first embodiment of the present invention. 1 is a microphone for inputting voice, 2 is an acoustic analysis unit that calculates a power spectrum from an input signal and outputs the total power of a voice band, 3 is a voice section detection unit that detects a voice section using voice band power, 4 is Noise spectrum estimating means for estimating a noise spectrum from data of a non-speech section, 5 a noise removing means for removing a noise component from data of a speech section, 6 a word feature vector extracting means for extracting a feature vector from the removed speech data. , 7 is a registered word dictionary storing the feature vectors of the registered words to be recognized, and 8 is a speech recognition means for comparing the extracted feature vectors with the registered words, calculating the likelihood thereof, and outputting the top word as a recognition result. . FIG. 2 shows an installation state of the voice recognition device, 1 is a microphone mounted on a headset, 30 is a voice recognition device according to the present invention, 40 is a controller such as a car navigation operated by the voice recognition device, and 50 is a car navigation system. Etc. are shown.

FIG. 3 shows a flow of processing of the acoustic analysis means 2. A preprocessing 21 performs windowing processing of the input signal, and a power spectrum calculation 22 calculates a power spectrum of the input signal. In the calculated power spectrum of the input signal, the total power of the audio band divided into the low band, the middle band, and the high band is integrated.

The noise power of the input signal is suppressed by integrating only the power of the voice band in the acoustic analysis means 2, and the voice section can be easily detected even under high noise. FIG. 4 shows an example of a case where “Sapporo” has occurred, and shows a method of determining voice section detection in the voice section detection means 3.
The speech band power calculated by the sound analysis means 2 is input to the speech section detection means 3 as input power. The noise level is sequentially updated in the section determined to be a non-voice section, and when the input power becomes larger than the threshold value 1, a temporary start end T1 is detected. When T1 is detected, the noise level is fixed at the last updated noise level value. If the input power becomes larger than the threshold value 2 during several frames after T1 is detected, T1 is set as the starting point Ts. When the input power becomes smaller than the threshold value 3 after detecting the start end Ts and further becomes smaller than the threshold value 4, the provisional end T2 is detected. If the input power does not exceed the threshold value 4 for a certain period of time (Tp) after detecting T2, a voice section is detected with T2 as the terminal Te. At this time, if the input power again exceeds the threshold value 4 within Tp, the provisional termination T2 is reset, and the provisional termination is detected again.

When the voice section detecting means 3 determines that the voice section is a non-voice section, the power spectrum of the input signal is supplied to the noise spectrum estimating means 4. Noise spectrum estimation means 4
Then, the power value is estimated for each frequency component and updated sequentially. As the estimation method, linear prediction, least square linear prediction, or the like can be used. When the voice section detecting means 3 detects the temporary start end T1, the noise removing means 5 removes the noise spectrum estimated by the noise spectrum estimating means 4 from the input signal power spectrum, thereby suppressing the noise component contained in the input signal. I do. Next, the word feature vector extraction unit 5 extracts a word feature vector from the power spectrum data from which the noise spectrum has been removed by the noise removal unit 5. The registered word dictionary 7 stores feature vectors of registered words. When the voice section is determined, the recognition processing means 8
The likelihood between the extracted word feature vector and the registered word feature vector is calculated. Here, when the likelihood difference between the top word and the next word is less than the threshold value R1, a signal prompting re-listening is output as a recognition result. If the likelihood difference between the top word and the next word is equal to or greater than the threshold value R2, the top word is output as a recognition result.

Embodiment 2 FIG. 5 shows a second embodiment of the present invention. Reference numeral 2 denotes an acoustic analysis unit that calculates a power spectrum from an input signal and outputs a total power of a voice band. 3 denotes a voice section detection unit that detects a voice section using the voice band power. 4 denotes a noise spectrum from data of a non-voice section. A noise spectrum estimating means for estimating the noise component; a noise removing means for removing a noise component from the data of the voice section;
Is a word feature vector extraction means for extracting a feature vector from the removed speech data, 7 is a registered word dictionary storing feature vectors of registered words to be recognized, 8 is a comparison of the extracted feature vector with the registered words and the likelihood thereof. Is a voice recognition means for calculating the top several words as a recognition result, and 9 is a keyword determination means for determining the keyword of the voice switch from the outputted top several words. Here, for example, a nickname such as “navi” may be set as the keyword.

The keyword determining means 9 realizes a voice switch. If the highest word among the top several words output from the voice recognizing means 8 is a keyword and the next word is a specific word, it is a keyword. Is determined. If the keyword is not included in the top few words output from the voice recognition means 8, the input word is discarded.
If the top word is a keyword and the next word is not a specific word, a signal prompting re-listening is output as a recognition result. Here, when it is determined that the keyword is the keyword, the keyword determination means 9 is turned on and outputs the keyword as a recognition result. Further, a recognition result is output for the next input voice. If there is no voice input for several seconds or more, the switch is turned off and the keyword is determined again. That is, after the keyword is determined, the operation is performed so as to recognize the voice command to be input next, and when the unvoiced portion occurs for a certain time, the operation is performed so as to perform the keyword determination again.

Embodiment 3 FIG. 6 shows a third embodiment of the present invention. Reference numeral 2 denotes an acoustic analysis unit that calculates a power spectrum from an input signal and outputs a total power of a voice band. 3 denotes a voice section detection unit that detects a voice section using the voice band power. 4 denotes a noise spectrum from data of a non-voice section. A noise spectrum estimating means for estimating the noise component; a noise removing means for removing a noise component from the data of the voice section;
Is a word feature vector extracting means for extracting a feature vector from the removed speech data, 7 is a registered word dictionary storing feature vectors of registered words to be recognized, 8 is a comparison of the extracted feature vector with the registered words and the likelihood thereof. Speech recognizing means for calculating the top several words as a recognition result, and SNR judgment 10 for controlling the recognition processing according to the SN ratio from the average power of the speech section frame and the average power of the noise frame detected by the speech section detecting means. Means.

When the voice section is detected by the voice section detection means 3, the SNR determination means 10 calculates the average power of all frames in the voice section. Next, when the ratio with the noise power in the non-speech section estimated by the noise spectrum estimating means 3 is equal to or larger than a predetermined threshold value, processing after noise removal is executed. If the value is less than a certain set threshold, the input voice is output as unclear as an SNR determination result, and the input signal is discarded.

The SNR determination means 10 is, of course, used in the second embodiment.
Can be added to the speech recognition device shown in FIG.

Embodiment 4 FIG. 7 shows a fourth embodiment of the present invention. In the first embodiment and the second embodiment, the voice band power is used to detect the voice section. In the fourth embodiment, the noise spectrum is obtained by the main frequency detection 24 from the noise spectrum obtained by the noise spectrum estimating means 4. A main frequency component having a large power is detected. In the audio band power calculation 23, the audio band power is calculated excluding the power near the main frequency of the detected noise.

Embodiment 5 FIG. 8 shows a fifth embodiment of the present invention. In the fifth embodiment, in the noise removal processing 25, a power spectrum is calculated by subtracting the noise spectrum acquired by the noise spectrum estimating means 3 from the power spectrum of the input signal. The voice band power is calculated from the calculated power spectrum.

[0022]

According to the first aspect of the present invention, the noise level in the voice band power is used as a reference in voice section detection using only one microphone, and the noise level is updated successively in the non-voice section. , The voice section can be correctly detected. In addition, the noise spectrum is estimated from the environmental noise signal mixed in the non-voice section, and the noise component is removed by subtracting the estimated noise spectrum from the input signal in the detected voice section. Stable speech recognition is possible even in a low S / N environment.

According to the second aspect of the present invention, the keyword to be a voice switch is determined by using the recognized words of the top several words. Since no switch operation is required, operability can be improved.

Further, according to the third invention, the S of the input voice is
If the N ratio is too low, the input is rejected without performing the recognition process, and thus there is an effect of suppressing erroneous recognition.

Further, according to the fourth and fifth aspects of the present invention, the configuration is made so as not to be easily influenced by environmental noise, so that a high S / N ratio for voice section detection can be ensured, and the effect of correctly detecting voice sections can be obtained. is there.

[Brief description of the drawings]

FIG. 1 is a first embodiment of a speech recognition apparatus according to the present invention;
FIG.

FIG. 2 is a diagram showing an installation state of a voice recognition device according to the present invention.

FIG. 3 is a diagram showing acoustic analysis means of the speech recognition device according to the present invention.

FIG. 4 is a diagram showing a voice section detecting means of the voice recognition device according to the present invention.

FIG. 5 is a second embodiment of the speech recognition apparatus according to the present invention.
FIG.

FIG. 6 is a third embodiment of the speech recognition apparatus according to the present invention.
FIG.

FIG. 7 is a fourth embodiment of the speech recognition apparatus according to the present invention;
FIG.

FIG. 8 is a fifth embodiment of the speech recognition apparatus according to the present invention.
FIG.

FIG. 9 is a diagram showing a configuration of a conventional voice recognition device.

FIG. 10 is a diagram showing an installation state of a conventional voice recognition device.

[Explanation of symbols]

Reference Signs List 1 microphone for sound, 1b microphone for noise, 2 sound analyzing means, 3 sound section detecting means, 4 noise spectrum estimating means, 5 noise removing means, 6 word feature vector extracting means, 7 word registration dictionary, 8 voice recognizing means, 9 Keyword determination means, 10 SNR determination means, 11 trigger button, 21 preprocessing, 22 power spectrum calculation, 2
3 Voice band power calculation, 24 main frequency detection, 25
Noise removal processing, 30 voice recognition device, 40 car navigation controller, 50 car navigation monitor.

Claims (5)

[Claims] 1. A microphone for inputting a voice, a sound analysis unit for calculating a power spectrum from an input signal and outputting a total power of a voice band, a voice section detecting unit for detecting a voice section using the voice band power, Noise spectrum estimating means for estimating a noise spectrum from non-speech section data, noise removing means for removing a noise component from speech section data, word feature vector extracting means for extracting a feature vector from the removed speech data, A registered word dictionary that stores the feature vectors of the registered words to be recognized, and the extracted feature vectors are compared with the registered words to calculate the likelihood, and the likelihood difference between the top word and the next candidate is equal to or greater than the set threshold. And a speech recognition means for outputting the result as a recognition result. 2. A microphone for inputting voice, an acoustic analysis unit for calculating a power spectrum from an input signal and outputting a total power of a voice band, a voice section detecting unit for detecting a voice section using the voice band power, Noise spectrum estimating means for estimating a noise spectrum from non-speech section data, noise removing means for removing a noise component from speech section data, word feature vector extracting means for extracting a feature vector from the removed speech data, A registered word dictionary that stores feature vectors of registered words to be recognized, a speech recognition unit that compares the extracted feature vectors with the registered word dictionary, calculates the likelihood thereof, and outputs the top few words as a recognition result,
A speech recognition apparatus, comprising: a keyword determination unit that determines a keyword of a voice switch using a sequence of top several words of a recognition result. 3. The method according to claim 1, wherein the average power of the voice section frame and the average power of the noise frame detected by the voice section detection means are used as S
3. The speech recognition device according to claim 1, further comprising an SNR determination unit for controlling a recognition process according to the N ratio. 4. A voice section is calculated by calculating a main frequency from an estimated noise spectrum and using a sum of powers of a power spectrum of input data excluding a vicinity of a main frequency component of noise. Or the speech recognition device according to 2. 5. A voice section is detected by using a sum of powers of voice bands in a power spectrum obtained from a difference between a power spectrum of input data and a power spectrum of estimated noise. A speech recognition device according to claim 1.