JPH06208393A - Voice recognizing device - Google Patents

Abstract

PURPOSE:To recover a voice level which is extremely subtracted by subtraction processing in a sound recognizing device having a noise suppression circuit. CONSTITUTION:A spectrum edit processing section 402 performs spectrum edit processing for an output of a subtraction section 404. Also, a noise level from a noise level measuring section 403 is inputted to a compensation quantity setting section 500, and the setting section 500 sets that how much % of the noise level is to be added to analysis data and returned. Set compensation quantity is inputted to a compensation switch section 502, and added to a compensation quantity addition section 503 in accordance with a judged result of a zero judging section 501. The zero judging section 501 judges whether it is zero or non-zero for data of after spectrum edit processing. At the time, if data is zero, the compensation switch section 502 is controlled so that compensation quantity is not inputted to the compensation quantity addition section 503. And if data is non-zero, since it can be judged as a voice section, the zero setting section 501 controls the compensation switch section 502 so that compensation quantity is inputted to the compensation quantity addition section 503, in order to compensate.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device,
In particular, it relates to a voice recognition device having a noise suppression circuit.

[0002]

2. Description of the Related Art In recent years, research and development of a system for controlling a device using a voice recognition device have been carried out and are being put to practical use. In particular, the operation of equipment in the interior of a car
By using the voice recognition device, hands can be used (hands-free) and the viewpoint can be kept (eyes-free), so that there is an advantage that driving safety in device operation can be secured and improved. However, there is a problem with the current voice recognition device (especially, such as vehicle running noise) that the voice recognition performance is significantly deteriorated in a noisy environment.

FIG. 8 is a block diagram showing a configuration of a conventional voice recognition device. The voice recognition device is roughly divided into a voice analysis unit 100, a voice recognition unit 200 and a dictionary unit 3.
It is composed of 00. In FIG. 8, the microphone 105
The voice analysis unit 100 performs feature extraction on the voice input from.

The voice analysis unit 100 has N-channel bandpass filters, each of which has an absolute value circuit (RECT).
And a low pass filter (LPF) is connected. The analysis data is sequentially input to the A / D converter and becomes digital data. Then, the voice analysis result (digital data) is output to the voice recognition unit 200 via the external bus 110.

The reason why the voice recognition performance of the conventional voice recognition apparatus as shown in FIG. 8 is deteriorated in a noisy environment is that the mixing of noise mainly transforms the spectrum pattern of the voice so that the dictionary is previously prepared. It can be mentioned that the similarity with the standard pattern registered in 300 is reduced.

Further, in the case of a word voice recognition device, it is known that a voice section detection process for detecting the start and end of a word voice that is sounded in isolation and cutting out a voice signal is required. If the sound is deformed or buried in noise, accurate voice detection cannot be performed, and the detection accuracy decreases, which leads to deterioration in recognition performance.

Further, when the ambient noise becomes high, the voice section cannot be detected, the recognition device side becomes inoperable, and the recognition result is not output even if the voice is uttered.

[0008] Therefore, as a measure against ambient noise, conventionally, the distance between the microphone and the speaker's mouth is reduced (to about 5 to 10 cm), and the microphone sensitivity is reduced so that the sound other than the speaker can be input from the microphone. In many cases, measures are taken to make it harder to be done.

However, in order to shorten the microphone distance, the method of installing the microphone becomes a problem. For example, the speaker wears the microphone, or the microphone is installed in front of the speaker's eyes. In such a case, for example, when the voice is recognized while driving a car, it is difficult to use if the microphone is in front of the eyes, etc. When it is used, the microphone has to be attached and detached each time the user gets in and out of the car, which causes an inconvenience such as annoying vocalization.

For this reason, when used in a vehicle interior, the microphone distance is often set to several tens of cm (about 30 to 50 cm) and attached to a sun visor or the like. According to this method, the inconvenience caused by the microphone mounting method is eliminated, but it is necessary to increase the microphone sensitivity as the microphone distance increases, and noise tends to be picked up from the microphone. Instead, it is necessary to take noise countermeasures on the side of the voice recognition device.

Since the running noise has a power biased to a low frequency, a method of relatively reducing a part of the noise by using an HPF (high-pass filter) as a preprocessing of the voice recognition device can be considered. However, in this method, the cutoff frequency must be appropriately selected because the voice information is also lost by using the HPF. It should be noted that in this case, the use of the HPF significantly reduces the low-frequency noise as compared with the loss of the voice information, so the S / N is relatively improved.

Countermeasures against running noise by HPF are shown, for example, in "Noise-Resistant Speech Recognition System" by Hamada and Takizawa (Matsushita Electric), Communication Techniques SP89-105. As a drawback of the above noise resistant voice recognition system, if the cutoff frequency is set to be high in order to significantly reduce noise, the loss of voice information also increases and the recognition performance decreases, so there is a limit to the setting of the cutoff frequency.
Therefore, there is a limit to the noise reduction, and noise exists in the band other than the noise component suppressed by the HPF, but no noise countermeasures have been taken for this part, so that the recognition performance is dramatically improved. However, this system is only one of the practical methods used as pretreatment.

As described above, the analysis data of the voice used for the voice recognition is greatly deformed due to the inclusion of noise, which causes the reduction of the recognition performance. And HP
There is a problem that the suppression of the noise component existing in the analysis band of the voice cannot be suppressed without the loss of the voice information only by suppressing the low frequency noise by F (or even without the HPF).

[0014]

As described above, the voice analysis data by the voice analysis unit 100 is output to the voice recognition unit 200. As a solution to the above problems, the voice analysis unit 100 and the voice recognition unit 200 are provided. A voice with improved noise resistance performance (voice recognition rate) by providing a noise countermeasure processing unit between them and performing waveform shaping (hereinafter referred to as "Specturam-Edit") on the analysis data of the voice deformed by noise. A recognizer has been proposed.

<Proposed Speech Recognition Device> FIG.
It is a block diagram which shows the structural example of the said speech recognition apparatus. In FIG. 9, the environmental noise subtraction unit 401 and the voice waveform shaping unit 402 are shown as the noise countermeasure processing unit 400, but the voice waveform shaping unit 402 may or may not be included in the noise countermeasure processing unit 400. Yes, the processing is independent of the environmental noise subtraction unit 401.

FIG. 3 shows the microphone 105 of the voice recognition apparatus of FIG.
It shows an example of an analysis result of the number "7" of channels of the voice analysis unit 100 of the voice "Sapporo" including the noise input from (hereinafter, CH1 to CH7 are channels (numbers) in FIGS. 3 to 7). Indicates).

In the apparatus of FIG. 8, the average noise level is measured by the environmental noise subtraction unit 401 for the analysis data shown in FIG. 3 and the noise component is subtracted from the original data to obtain spectrum subtraction. I was going. FIG. 4 is an output waveform diagram for each channel of the sound "Sapporo" according to the result of the spectral subtraction.

The voice waveform shaping section 402 of FIG. 9 detects a range in which no voice exists or a silent section of the voice and replaces that portion with zero data (spectrum edit processing). FIG. 5 is a waveform diagram for each channel of the sound "Sapporo" by the spectrum edit processing.

However, in the speech recognition apparatus proposed above, when the average noise level is high, the speech level after the subtraction processing becomes lower than the level that would have existed when there was no noise. Inconvenience occurs in section detection. The reason is that the speech segment detection algorithm of the word speech recognition device is generally configured to operate according to a parameter determined from cut-and-try and many data,
It can be mentioned that the operation of detecting the voice section is remarkably dependent on the appearance of the peaks and valleys of the voice pattern.

From this, in order to accurately detect the voice section, subtraction is performed.
It is desirable that the shape and level of peaks and valleys of the voice pattern do not change due to the processing. In other words, it is ideal that the noise is suppressed in the data after the subtraction, and the ultimate problem is that the voice remaining after the processing is the same as the voice without noise.

However, if the noise level is high, the sound is drawn (reduced) too much and the level becomes lower than the original voice as described above, the voice section detection operation becomes unstable, and the similarity of the voice pattern to the dictionary is low. Will decrease. At this time, if the subtraction amount is set to a value smaller than the average noise level in order to prevent the sound level from decreasing, there is a disadvantage that the noise cannot be suppressed sufficiently.

FIGS. 6 and 7 are examples of such a case, FIG. 6 is a waveform of voice data after subtraction, and FIG. 7 is a waveform of analysis data of voice “Sapporo” in the absence of noise. 6 and the waveforms of FIG. 7 are compared, CH of FIG.
It can be seen that the amplitude of the data of (channel) 1 is small. This is because the input noise is a simulation of running noise, and its power is concentrated in the low range, so the noise level included in channel 1, which is the low range channel, becomes high, and as a result, subtraction is performed during subtraction. Due to the large amount.

The present invention has been made in view of the above inconvenience, and an object of the present invention is to provide a voice recognition device capable of restoring a voice level that has been excessively subtracted by a subtraction process in a voice recognition device having a noise suppression circuit. .

[0024]

In order to achieve the above object, a voice recognition apparatus according to the first invention comprises a voice analysis means for converting a voice captured by a voice input means into a first voice analysis data. Based on the noise detecting means for detecting the noise component from the voice analysis data to obtain a noise component signal, the subtracting means for obtaining the second voice analysis data by subtracting the noise component from the voice analysis data, and the second voice analysis data, A voice recognition device comprising: a voice recognition means for recognizing voice to obtain a voice recognition result; a voice section detecting means for detecting a voice section of the second voice analysis data to obtain a detection signal; and a noise component signal location. Correction processing means for giving a fixed amount to the second voice analysis data and correcting the second voice analysis data;
It is characterized by having.

A second invention is the speech recognition apparatus according to the first invention, further comprising a spectrum edit means for subjecting the second speech analysis data to a spectrum edit process to obtain a spectrum edit signal, and a voice section analysis means. Is a zero section detecting means for obtaining a detection signal by detecting a zero section or a non-zero section of the spectrum edit signal, the correction processing means, a correction amount setting means for obtaining a predetermined amount of the noise component signal from the noise component signal, It is characterized in that it has a synthesizing means for synthesizing the spectrum edit signal and the noise component signal of a predetermined amount, and a control means for controlling the synthesizing processing based on the detection signal from the zero section detecting means.

A third invention is the speech recognition apparatus according to the first invention, wherein the correction processing means obtains a predetermined amount of the noise component signal from the noise component signal, and a second correction amount setting means.
And a control means for controlling the synthesizing process based on the detection signal from the voice section detecting means.

[0027]

With the above arrangement, the voice recognition apparatus according to the first aspect of the invention detects the voice section of the second voice analysis data by the voice section detecting means to obtain a detection signal, and the correction processing means to determine a predetermined amount of the noise component signal. It is given to the second voice analysis data, and the correction processing of the second voice analysis data is performed.

A second aspect of the present invention is the voice recognition apparatus according to the first aspect of the present invention, further comprising: spectrum-editing processing the second voice-analysis data to obtain a spectrum-editing signal. The zero section or non-zero section of the spectrum edit signal is detected to obtain a detection signal. Then, the correction processing means obtains a predetermined amount of the noise component signal from the noise component signal by the correction amount setting means, and performs the combining processing of the spectrum edit signal and the predetermined amount of noise component signal by the combining means,
The control means controls the synthesizing process based on the detection signal from the zero section detecting means.

According to a third invention, in the speech recognition apparatus according to the first invention, the correction processing means obtains a predetermined amount of the noise component signal from the noise component signal by the correction amount setting means,
The synthesizing unit performs the synthesizing process of the second voice analysis data and the noise component signal of the predetermined amount, and the controlling unit controls the synthesizing process based on the detection signal from the voice section detecting unit.

[0030]

【Example】

<Embodiment 1> FIG. 1 is a block diagram showing an embodiment of a voice recognition device according to the first invention, and 400 is a voice suppression device provided between a voice analysis unit 100 and a voice recognition unit 200. The voice suppression device 400 includes the environmental noise subtraction unit 40.
1, a spectrum edit processing unit 402, a correction amount setting unit 500, a zero determination unit 501, a correction switch unit 502, and a correction amount addition unit 503. Also, the environmental noise subtraction unit 401 includes a noise level measurement unit 403 and a subtraction unit 404.
It is composed of.

A voice containing noise (for example, refer to the waveform of the voice "Sapporo" in FIG. 3) input from the microphone 105 to the voice analysis unit 100 is analyzed by the voice analysis unit 100 and output as voice analysis data (digital data). And input to the environmental noise subtraction unit 401 for each channel. The noise level measuring unit 403 measures the average noise level γ for each channel.

Further, the subtracting section 404 performs a process of subtracting the measured noise level from the original voice analysis data to obtain an output waveform for each channel as shown in FIG. The spectrum edit processing section 402 performs spectrum edit processing on the output of the subtraction section 404 to obtain an output waveform for each channel as shown in FIG. Then, the correction amount setting unit 500 receives the noise level γ from the noise level measuring unit 403.
Is input, and what percentage of the noise level is added back to the analysis data as shown in FIG. 5 is set. Here, this setting rate is α
Then, the correction amount to be added back is αγ.

The correction amount set by the correction amount setting unit 500 is input to the correction switch unit 502 and added to the correction amount addition unit 503 according to the determination result of the zero determination unit 501,
It is not done.

The zero decision unit 501 decides whether the data (FIG. 5) after the spectrum edit processing is zero or non-zero. At this time, if the data is zero, the correction amount αγ is corrected so that the correction amount addition unit 503 does not input the correction amount αγ.
Control 02. Further, when the data is non-zero, it can be determined that there is a voice section, and therefore the zero determination unit 501 controls the correction switch 502 so that the correction amount αγ is input to the correction addition unit 503.

FIG. 7 shows a case where the voice analysis data shown in FIG. 5 is subjected to correction processing, and is an example (α = 0.5) in which ½ of the subtraction amount is added back. In this case, since the subtraction amount is large in the channel with a high noise level (CH1 in the embodiment), the correction amount is also large. In addition, in CH1 of FIG. 7, the section and the section are non-zero, and are the sections to which the correction amount is added. It is clear from FIG. 7 that the audio level has recovered.

<Embodiment 2> FIG. 2 is a block diagram showing an embodiment of the speech recognition apparatus according to the first invention, in which the waveform after subtraction is directly corrected without performing spectrum edit processing. . In FIG. 2, the voice suppression device 400 includes an environmental noise subtraction unit 401, a correction amount setting unit 500, a voice section detection processing unit 504, a correction switch 502, and a correction amount addition unit 5.
It is composed of 03. Also, the environmental noise subtraction unit 401
Is the same as that of FIG.

In FIG. 2, from the microphone to the voice analysis unit 10
The voice including noise input to 0 (for example, refer to the waveform of the voice “Sapporo” in FIG. 3) is analyzed by the voice analysis unit 100, output as voice analysis data (digital data), and is channeled to the environmental noise subtraction unit 401. It is input every time. The noise level measuring unit 403 measures the average noise level γ for each channel. Further, the subtracting unit 404 performs a process of subtracting the measured noise level from the original voice analysis data to obtain an output waveform for each channel as shown in FIG.

If the spectrum edit process is not performed, it is not possible to specify the section to be corrected by the zero / non-zero judgment because there is noise that cannot be completely subtracted and remains in addition to the voice section to be corrected. Therefore, the noise section detection unit 504
The voice section is detected by. On the other hand, the correction amount setting unit 500
1 inputs the noise level γ from the noise level measuring unit 403 as in the case of FIG. 1 and sets what percentage of the noise level is added back to the analysis data as shown in FIG. Here, when the set rate is α, the correction amount to be added back is αγ.

The correction amount set by the correction amount setting unit 500 is input to the correction switch unit 502, and the voice section detection unit 50
Correction amount adding section 50 according to the result of voice section detection by
It may or may not be added to 3.

[0040]

As described above, according to the first aspect of the present invention, the voice level lost by the spectral subtraction (subtraction process) can be restored, and the waveform after spectrum editing (waveform shaping) can be input first. It becomes possible to approximate the waveform of the voice, and the similarity of the waveform to the dictionary is improved, so that the recognition rate in the voice recognition unit can be improved. Further, according to the second invention, it is possible to restore the voice level lost by the spectrum subtraction without performing the spectrum edit process,
With a simpler circuit configuration, the similarity of the waveform to the dictionary is improved, so that the recognition rate in the voice recognition unit can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a voice recognition device according to the first invention.

FIG. 2 is a block diagram showing an embodiment of a voice recognition device according to the first invention.

FIG. 3 is an example of an analysis result by a voice analysis unit of a voice “Sapporo” including noise.

FIG. 4 is a waveform diagram of a voice “Sapporo” showing a result of spectrum subtraction.

FIG. 5 is a waveform diagram of the sound “Sapporo” showing the result of the spectrum editing process.

FIG. 6 is voice analysis data of voice “Sapporo” when there is no noise.

7 is a waveform chart showing a result of a correction process performed on the voice analysis data shown in FIG.

FIG. 8 is a block diagram showing a configuration example of a conventional voice recognition device.

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

[Explanation of symbols]

100 voice analysis unit (voice analysis unit) 105 microphone (voice input unit) 200 voice recognition unit (voice recognition unit) 402 spectrum edit processing unit (spectrum edit unit) 403 noise level measurement unit (noise detection unit) 404 subtraction unit (subtraction) Means) 500 correction amount setting unit (correction amount setting unit) 501 zero determination unit (zero section detection unit) 502 correction switch unit (control unit) 503 correction amount addition unit (combining unit)

Claims (3)

[Claims] 1. A voice analysis unit for converting a voice captured by a voice input unit into first voice analysis data, a noise detection unit for detecting a noise component from the voice analysis data to obtain a noise component signal, and the voice. And a subtraction unit for obtaining second voice analysis data obtained by subtracting the noise component from the analysis data, and a voice recognition unit for recognizing the voice and obtaining a voice recognition result based on the second voice analysis data. In the voice recognition device, a voice section detecting means for detecting a voice section of the second voice analysis data to obtain a detection signal, and a predetermined amount of the noise component signal is given to the second voice analysis data to obtain the second voice analysis data. A voice recognition device, comprising: a correction processing unit that performs correction processing of voice analysis data. 2. The voice recognition device according to claim 1, wherein
Further, the second voice analysis data has a spectrum edit means for performing spectrum edit processing to obtain a spectrum edit signal, and the voice section analysis means detects a zero section or a non-zero section of the spectrum edit signal to obtain a detection signal. Correction section setting means for obtaining a predetermined amount of the noise component signal from the noise component signal, and synthesizing means for synthesizing the spectrum edit signal and the noise component signal of the predetermined amount. And a control unit that controls the synthesis process based on a detection signal from the zero section detection unit. 3. The voice recognition apparatus according to claim 1, wherein the correction processing means includes a correction amount setting means for obtaining a predetermined amount of the noise component signal from the noise component signal, the second voice analysis data and the predetermined amount. A voice recognition device comprising: a synthesizing unit that performs a synthesizing process with a noise component signal; and a control unit that controls the synthesizing process based on a detection signal from a voice section detecting unit.