JP2001296887A - Speech recognition method and speech recognition device using the method - Google Patents

Abstract

(57) [Summary] [PROBLEMS] To provide a speech recognition method and apparatus for improving the recognition rate under periodic noise. In a voice recognition method under noise, a cutoff frequency determination unit and a high-pass filter controlled by the cutoff frequency determination unit are provided downstream of a voice detection unit. The cutoff frequency determination unit 30 analyzes the main components of noise in each band of the audio signal. If the main component of the noise is periodic noise, the band is set in the high-pass filter 40 as a cutoff frequency.
Thus, even if the band of the periodic noise changes, it is always reduced. Therefore, the speech recognition rate is improved. If random noise is dominant, a predetermined cutoff frequency is set. As a result, the human voice band is not cut and the recognition rate is not reduced. Thereby, a high speech recognition rate can be always maintained.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method for improving a speech recognition rate under noise and a speech recognition apparatus using the method. In particular, the present invention relates to a voice recognition method and a voice recognition device that increase the recognition rate of human voice in an environment where periodic noise exists. INDUSTRIAL APPLICABILITY The present invention can be applied to a vehicle interior voice recognition device in which periodic noise changes depending on a vehicle type or a running state of a vehicle.

[0002]

2. Description of the Related Art Conventionally, there is a speech recognition method and apparatus for recognizing human speech under noise. For example, there is a development of a noise-tolerant speech recognition device (1) (Proceedings of the Acoustical Society of Japan, March 1989, p. 117). This is a speech recognition method under traffic noise and domestic noise. Since the noise of the vehicle is concentrated in the low frequency band, a high-pass filter for suppressing the low frequency band is usually used. In the above paper, the high-pass filter is used as preprocessing for the speech recognition processing, and the relationship between the band to be suppressed and the recognition rate is examined. The result of the recognition indicates that the cutoff frequency is effective at about 300 Hz on average.
In addition, there is a speech recognition apparatus and method disclosed in Japanese Patent Application Laid-Open No. 11-231897. This is a method and apparatus for performing voice recognition by removing low-frequency noise immediately before a voice section in a frequency space for human voice detection.

[0003]

However, noise generally includes stationary noise (random noise) and periodic noise.
Speech recognition may be needed under periodic noise. An important vowel in human voice recognition is a voiced sound, and the voiced sound is a complex periodic sound having a fundamental frequency in a band of 100 to 300 Hz. For example, when recognizing a driver's voice in a vehicle cabin, periodic noise caused by, for example, an engine sound of the vehicle is mixed in the band.

[0004] The development of a noise-tolerant speech recognition device (1) (Proceedings of the Acoustical Society of Japan, March 1989, p117)
The cutoff frequency of the high-pass filter is fixed at 300 Hz. This means that a part of the human voice band is also cut, and in some cases, the recognition rate is reduced. this is,
This is also pointed out in Japanese Patent Application Laid-Open No. 06-208393. However, if the noise includes periodic noise having a large level, the recognition rate may be increased by increasing the cutoff frequency even if some audio information is cut. That is, a method of using a fixed high-pass filter as preprocessing for speech recognition is not necessarily a recognition method for improving the recognition rate. Therefore, an apparatus using the method has not always been a speech recognition apparatus that improves the recognition rate.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problems, and has as its object to focus on the fact that the band of the actual periodic noise changes, and to cut the high-pass filter according to the band. An object of the present invention is to improve the speech recognition rate under periodic noise by adjusting an off-frequency and extracting an optimal speech. It is another object of the present invention to apply the voice recognition method of the present invention to a voice recognition method in a vehicle cabin so that a driver's voice can be clearly recognized to improve driving safety.

[0006]

According to the first aspect of the present invention, there is provided a voice recognition method under noise, wherein a periodic sound component forming each frequency band of voice including noise is extracted, and a periodic sound component is extracted. Is characterized by performing a speech recognition by suppressing a frequency band of which is a subject by a filter.

Further, the voice recognition method according to the present invention is characterized in that the filter is a high-pass filter, and the cut-off frequency of the high-pass filter is adjusted according to a frequency band mainly composed of a periodic sound component. .

According to the third aspect of the present invention, the noise is noise in a vehicle interior, and the cutoff frequency of the high-pass filter is a frequency band for a periodic sound component caused by a vehicle type or a running state of the vehicle. It is characterized by being adjusted according to.

According to a fourth aspect of the present invention, in the voice recognition apparatus for recognizing voice under noise, voice detection means for detecting voice, and periodic sound of each frequency band based on an output of the voice detection means. Cut-off frequency determining means for extracting a component and determining a cut-off frequency according to a frequency band mainly composed of a periodic sound component, a high-pass filter for setting the cut-off frequency, and a sound obtained by the high-pass filter And a noise recognizing means for recognizing and processing the voice processed by the noise suppressing means.

[0010]

According to the speech recognition method of the present invention, first, a periodic sound component forming each frequency band is extracted from speech including noise. Then, a frequency band mainly composed of the periodic sound component is specified, and the band is suppressed by the filter. This means that the periodic noise is removed from the speech to be recognized. Then, voice recognition is performed. Since the human voice is not always present in the specific band, the filter may partially suppress, but not completely suppress the sound in some cases. That is,
Only periodic noise is suppressed, and human voice is not suppressed. That is, only the human voice is relatively extracted. Therefore, by using the method of the present invention, background noise can be effectively removed, and the speech recognition rate can be improved.

According to the speech recognition method of the present invention, the filter is a high-pass filter, and the cut-off frequency of the high-pass filter is adjusted according to a frequency band mainly composed of a periodic sound component. For example, periodic noise composed of motor sounds exists in a band having a higher frequency than human voice. Then, the periodic noise changes according to the rotation speed. For example, during high-speed rotation, the frequency shifts to a higher frequency. That is, the frequency band mainly composed of the periodic sound component changes. The cutoff frequency of the high-pass filter is adjusted according to the changing frequency band. That is, the high-pass filter can completely suppress the changing periodic noise. In other words, human voices can be extracted relatively efficiently. Therefore, this is a voice recognition method that always maintains a high voice recognition rate.

According to the third aspect of the present invention, the noise is the noise in the vehicle interior, and the cutoff frequency of the high-pass filter is determined according to the frequency band for the periodic sound component caused by the vehicle type or the running state of the vehicle. Has been adjusted. Generally, the noise in the vehicle cabin is a noise caused by an engine sound of the vehicle or a noise caused by running. Therefore,
The band of the noise changes with the traveling speed. In the present invention, the cutoff frequency of the high-pass filter is adjusted according to the frequency band change of noise caused by the type of vehicle or the running state of the vehicle (engine speed). In other words, the high-pass filter is set so as to always reduce them according to the noise change due to the vehicle type or the running state of the vehicle. In other words, the background noise is always removed, and only the voice of the driver is extracted. Therefore, by using the method of the present invention, a high speech recognition rate can be maintained even in a vehicle cabin.

Further, according to the voice recognition device of the fourth aspect, the voice detection means first detects voice including noise. Next, the cutoff frequency determining means extracts the periodic sound component of each frequency band from the output of the voice detecting means, and determines the frequency band mainly composed of the periodic sound component. Then, the cutoff frequency of the high-pass filter is controlled so as to suppress the next determined frequency band. That is, the cutoff frequency of the high-pass filter is adjusted according to the frequency band in which the periodic sound component is the main component. The high-pass filter suppresses the periodic sound component, that is, the noise in the frequency band where the level of the periodic noise is large, and then removes the noise by the noise suppression unit. The noise in this case is mainly random noise. For the removal of the noise, for example, a spectral subtraction method is used. Then, the speech processed by the noise suppression unit is recognized by the speech recognition unit.

In the speech recognition apparatus according to the present invention, the cutoff frequency of the high-pass filter is adjusted according to the frequency band of the periodic sound component (periodic noise) by the above-described method. Therefore, it is possible to cope with various periodic noises, in particular, periodic noises whose bands change. That is, it is possible to minimize the influence on the human voice and always effectively suppress only the periodic noise. Therefore, the speech recognition apparatus can improve the speech recognition rate even under periodic noise. When this device is used, for example, for voice recognition in a vehicle cabin, it is possible to accurately suppress periodic noise that changes depending on the vehicle type and the running state of the vehicle, and to reliably increase the voice recognition rate of the driver.

[0015]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A speech recognition method and apparatus according to the present invention will be described below with reference to the drawings. As described above, the speech recognition method and apparatus of the present invention are particularly effective under periodic noise. The periodic noise is a general noise having periodicity such as a traffic noise, a noise inside a moving body having an engine, and a mechanical sound at a factory or the like. Here, a speech recognition device under vehicle interior noise will be described.

The vehicle interior noise includes periodic noise caused by engine noise and random noise (road noise, wind noise, etc.) caused by running. These vehicle interior noises are concentrated in the low frequency band. Therefore, at the time of speech recognition, preprocessing for removing them using a high-pass filter is effective. However, at present, even if a high-pass filter is used, the recognition rate has not been further improved. Then, the inventor paid attention to the fact that the frequency band of the periodic noise changes depending on the vehicle type and the running state, and clarified that this is related to the recognition rate. The method is
Analyzes the acoustic characteristics (whether the main component of the periodic sound component) of the periodic noise that changes depending on the vehicle type and driving conditions, and based on the analysis, optimally controls the cutoff frequency of the high-pass filter, and performs speech recognition. The rate has been improved.

FIG. 1 shows an embodiment of a speech recognition apparatus according to the present invention. The figure is a system block diagram. The voice recognition device according to the present invention includes a microphone 1 that collects a driver's voice.
0, a voice detection unit 20 as a voice detection unit, a cutoff frequency determination unit 30 as a cutoff frequency determination unit, a high-pass filter 40, and a noise suppression unit 5 as a noise suppression unit
0, and a voice recognition unit 60 as voice recognition means.

Each component and its function will be described according to the flow of the audio signal. First, the voice of the driver is detected by the voice detection unit 20 via the microphone 10. The voice detection unit 20 detects the periodic noise and the random noise in addition to the human voice. Then, the voice section of the driver is detected from among them, and is sent to the cutoff frequency determination unit 30.

The cut-off frequency determining section 30 first analyzes the background noise and determines whether or not a periodic sound component is the main component for each frequency band. This determination is, for example, frequency spectrum analysis and spectrum envelope analysis. It is determined that a band having a large component protruding from the spectral envelope is mainly composed of a periodic sound component. The periodic sound component can be regarded as periodic noise caused by the engine sound. FIG. 2 (a),
(B) shows a noise spectrum diagram during acceleration running of the two types of vehicles. (A) is a vehicle type and (b) is a b vehicle type. The vehicle type a is determined to be mainly composed of periodic noise in the 200 to 400 Hz band, and the vehicle type b is determined to be mainly composed of random components in the band. The thin line is the frequency spectrum, and the thick line is the spectrum envelope.

From the above noise spectrum diagram, the cut-off frequency determining unit 30 determines that the high-pass filter 4
The cutoff frequency of 0 is determined to be 400 Hz. Then, it is output to the high-pass filter 40. The high-pass filter 40 cuts an audio signal of 400 Hz or less and outputs the signal to the noise suppression unit 50 at the next stage. That is, the periodic noise is suppressed and output. The noise suppression unit 50 is a noise suppression process that is generally performed. That is, for example, a spectrum subtraction process of subtracting a spectrum component immediately before a voice section from a spectrum of the voice section. Thereby, the background noise is removed and the voice of the driver is extracted. After that, the voice of the carrier is recognized by the normal voice recognition unit 60. Thereby, the recognition rate is improved.

FIG. 2A shows the increase / decrease of the recognition rate depending on the cutoff frequency of the high-pass filter 40. In the figure, H
PF means the cutoff frequency of the high-pass filter 40. As shown in the figure, the recognition rate when the cutoff frequency of the high-pass filter 40 is 400 Hz is about 57.6.
%, Which is about 33.7% in the case of the conventional 200 Hz. Conventionally, when the vehicle is accelerating, the engine sound component increases in the band near the fundamental frequency (100 to 300 Hz) of the voice and the recognition rate is greatly deteriorated (33.7%). Improving the speech recognition rate by setting the frequency to 400 Hz (57.6
%)We were able to.

On the other hand, it is determined that the random noise is dominant for the vehicle type b. In this case, since the periodic noise is not the main component, the cutoff frequency determination unit 30 determines the cutoff frequency to be a predetermined 200 Hz. Then, a series of processes described above are performed. The lower part of FIG. 2B shows the increase / decrease of the recognition rate depending on the cutoff frequency of the high-pass filter 40. When the cutoff frequency is set to 200 Hz, the recognition rate is about 67.4%, and when it is set to 400 Hz, the recognition rate is reduced to about 58.1%. In other words, the cutoff frequency is not set to 400 Hz and the
By setting z, a high recognition rate can be maintained.

As described above, when the main component of the noise is determined and the cutoff frequency of the high-pass filter is optimally adjusted based on the discrimination, the periodic noise can be surely reduced. That is, the use of the voice recognition method and apparatus of the present invention makes it possible to always clearly extract the driver's voice even in the interior of an automobile where the periodic noise changes, and to surely increase the voice recognition rate.

(Modifications) The above embodiment is one example, and various other modifications are possible. For example, in the above embodiment, a high-pass filter is used, but a band-pass filter or a low-pass filter may be used. A band where periodic noise exists may be cut by a band-pass filter or a low-pass filter. Similar effects can be obtained.

The speech recognition method and apparatus of the present invention are effective not only for the noise in the interior of a car but also for general noise including periodic sounds. It is effective for voice recognition devices in railways, ships, and aircraft. It is also effective for voice recognition in environments with periodic mechanical sounds, such as factories and construction sites.

[Brief description of the drawings]

FIG. 1 is a system block diagram of a speech recognition device according to an embodiment of the present invention.

FIG. 2 is a diagram illustrating a relationship between a noise spectrum, a cutoff frequency, and a recognition rate according to the embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 10 Microphone 20 Voice detection part 30 Cutoff frequency determination part 40 High-pass filter 50 Noise suppression part 60 Voice recognition part

Claims (4)

[Claims] 1. A method for recognizing speech under noise, comprising extracting a periodic sound component forming each frequency band of a sound including noise, suppressing a frequency band mainly composed of the periodic sound component by a filter, and A speech recognition method characterized by performing recognition. 2. The filter according to claim 1, wherein said filter is a high-pass filter.
The speech recognition method according to claim 1, wherein a cutoff frequency of the high-pass filter is adjusted according to a frequency band in which the periodic sound component is dominant. 3. The apparatus according to claim 2, wherein the noise is noise in a vehicle compartment, and a cutoff frequency of the high-pass filter is adjusted according to a frequency band for the periodic sound component caused by a vehicle type or a running state of the vehicle. The voice recognition method according to claim 2. 4. A speech recognition apparatus for recognizing speech under noise, comprising: speech detection means for detecting speech; and periodic sound components of each frequency band extracted from the output of the speech detection means. Cut-off frequency determining means for determining a cut-off frequency according to a certain frequency band; a high-pass filter for setting the cut-off frequency; a noise suppressing means for suppressing noise from voice obtained by the high-pass filter; A speech recognition unit for recognizing and processing the speech processed by the noise suppression unit.