JPH07191696A - Speech recognition device - Google Patents

Abstract

PURPOSE:To prevent misdetection due to the mixture of the feature quantity of an ambient noise with the feature quantity of a speech by reducing the influence of the ambient noise upon a voiceless section of a double consonant, etc., or a section wherein a speech level is low as to an inputted speech signal. CONSTITUTION:As for the inputted sound signal, acoustic feature quantities and the degrees of sound likelihood are found, frame by frame. At this time, when there is a decided section which the degree of sound likelihood is low, i.e., a section which is voiceless or low in speech level, the acoustic feature quantity of the section is made into a white noise by a noise processing part 6 and the spectrum representing the acoustic feature quantity is smoothed; and a matching process part 8 matches a time series of acoustic feature quantities after the process of the noise processing part 6 with a time series of a standard pattern to perform speech recognition. Therefore, the influence of an ambient noise on the section which is voiceless or low in speech level is removed to perform the speech recognition accurately.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition apparatus for recognizing an input voice by comparing the feature amount of the input voice with the feature amount of a standard pattern prepared in advance.

[0002]

2. Description of the Related Art In recent years, a voice recognition technique for recognizing phonological information of a language generated by humans has been actively researched, and a voice recognition device in which the voice recognition technique is applied to a concrete device is being developed. . To recognize phonological information, in general,
Prepare a plurality of standard patterns in units of words, syllables, etc. in advance, compare the unknown input voice and each standard pattern, find the standard pattern that is most similar to the input voice,
Techniques such as determining that this standard pattern is uttered voice are used.

Techniques for recognizing words, syllables, and the like in such speech recognition techniques include isolated word speech recognition for recognizing words and the like generated by dividing them, and specific words and the like for continuously generating sounds. There are two types of recognition: continuous word speech recognition. In practical use of a voice recognition device using the voice recognition technology, it is desirable to use a device that can realize continuous word voice recognition in consideration of noise generated in the surroundings and unnecessary words that a speaker may generate.

Word spotting such as continuous DP method has been conventionally known as a technique of continuous word voice recognition for recognizing voice by excluding ambient noise and unnecessary words generated by a speaker. Sadahi's "Digital Speech Processing"
Tokai University Press Chapter 8). Here, word spotting is a technique for searching a voice for a unit such as a word or a syllable and extracting a predetermined word. In addition, the continuous DP method (continuous Dynam
ic Programming) is a DP matching (Dynamic Programming matchi) for an input voice converted into a parameter sequence such as a spectrum by shifting the input voice one frame at a time from the start end and a standard pattern such as a word or a syllable.
ng), and when the distance as a result of matching falls below a certain threshold, it is a continuous word speech recognition that determines that a word or syllable of the standard pattern exists at that time.

Here, when comparing the input voice and the standard pattern in the voice recognition, the input voice is not compared with the voice waveform itself, but the phase information is removed from the voice waveform and converted into a feature related to the spectrum. It is common to handle information. This is because the amount of information is too large when comparing the voice waveforms themselves, and the phase information of the waveform is likely to change depending on the transmission system and the recording system, and such phase information hardly contributes to human perception of voice. Because.

As a characteristic relating to the spectrum, a short-time spectrum extracted at regular intervals is generally used. This short-time spectrum means the power spectrum density for each short-term section of speech, and the spectrum envelope that is a component that gently changes with frequency and the spectrum fine structure that is a component that finely changes with frequency. It can be analyzed by breaking it down into the product of (and sum on a logarithmic scale). Among these, the spectral fine structure is unstable under the influence of pitch and the like. Therefore, in speech recognition, the spectrum envelope is extracted from the short-time spectrum,
It is common practice to use the spectral envelope as a feature of speech.

There are various types of methods for extracting the spectral envelope, and one of them is the cepstrum analysis. The cepstrum analysis is defined as the inverse Fourier of the logarithm of the short-time amplitude spectrum of the waveform, and has a characteristic in that the spectrum envelope and the spectrum fine structure can be approximately separated. Further, as a method for extracting a spectrum envelope related to cepstrum analysis, an attempt has recently been made to use a cepstrum calculated from a logarithmic spectrum resampled at a mel-scale frequency. Such a cepstrum is called a mel cepstrum. Furthermore, as a special type of cepstrum analysis, LPC cepstrum analysis (LPC
Is a linear predictive coding: an abbreviation for linear prediction). The LPC cepstrum is a cepstrum calculated directly from a waveform, that is, an FFT cepstrum (FFT is a fast Fourier transfo
rm: an abbreviation for Fast Fourier Transform), means a cepstrum based on a linear prediction model, and is characterized in that an envelope spectrum in which the peak of the spectrum is emphasized rather than the envelope spectrum according to the FFT cepstrum can be obtained. In other words, focusing on the fact that important information for speech recognition exists in the peak part of the spectrum, by emphasizing the peak of the spectrum, we made the distance measure sensitive to realize more accurate speech recognition. It is a technique.

[0008]

As described above, for the detection of the speech feature amount, that is, the extraction of the spectrum envelope of the short time spectrum, a method such as cepstrum analysis, mel cepstrum analysis, or LPC cepstrum analysis is used. . At this time, a cepstrum coefficient, a mel cepstrum coefficient, and an LPC cepstrum coefficient are used as the audio feature amount, respectively. However, since such a cepstrum coefficient is a feature amount that does not depend on the input level of the voice, the feature amount of ambient noise is the feature amount of the input voice in a silent section such as when a phonation is made or a section where the voice level is low. However, there is a problem in that it may cause misrecognition. For example, in a silent section, the distance from the standard pattern of the word corresponding to the input voice is spread by surrounding noise, and it is mistaken for a word that does not correspond to the input pattern. There is a case where it is narrowed by the noise of and is mistakenly recognized as a corresponding word, which is an obstacle to realizing accurate voice recognition.

[0009]

The invention according to claim 1 is
An acoustic analysis unit that obtains the acoustic feature amount of the input acoustic signal by performing acoustic analysis for each frame, and a voice detection that obtains the degree of voice-likeness for each frame of the input acoustic signal. Section, a noise reduction processing section for converting the acoustic feature amount of the section determined to have a low degree of voice likeness into white noise by the voice detection section, a standard pattern storage section for storing a standard pattern of the voice, and the standard pattern. A matching processing unit for matching the time series of the standard pattern stored in the storage unit with the time series of the acoustic feature amount processed by the noise processing unit is provided.

According to a second aspect of the present invention, in the first aspect of the invention, the acoustic analysis unit performs a cepstrum analysis for obtaining a cepstrum coefficient as an acoustic feature amount for each frame, and the noise reduction processing unit obtains the cepstrum coefficient. By setting a small cepstrum coefficient of a frame having a low degree of speech likeness, the acoustic feature amount is converted to white noise.

According to a third aspect of the present invention, in the first aspect of the invention, the acoustic analysis unit performs a mel-cepstral analysis for obtaining a mel-cepstrum coefficient as an acoustic feature amount for each frame, and the noise reduction processing unit uses a speech detection unit. By setting the mel-cepstral coefficient of the frame having a low degree of voice-likeness to be small, the acoustic feature quantity is converted into white noise.

According to a fourth aspect of the present invention, in the first aspect of the present invention, the acoustic analysis unit performs a spectrum analysis to obtain a short-time spectrum for which correction is performed to remove a spectrum inclination as an acoustic feature amount for each frame, The noise conversion processing unit converts the acoustic feature amount into white noise by setting a small short-time spectrum of a frame having a low degree of voice likeness obtained by the voice detection unit.

According to a fifth aspect of the present invention, in the first aspect of the present invention, the voice detection unit determines that the smaller the power of the input acoustic signal is, the lower the likelihood of voice is, and the noise conversion processing unit determines that voice. As the detection unit determines that the degree of voice-likeness is low, the acoustic feature amount is strongly converted into white noise.

According to a sixth aspect of the present invention, in the first aspect of the invention, the standard pattern stored in the standard pattern storage section is subjected to processing equivalent to white noise conversion of the acoustic feature quantity in the noise processing section. This is the generated standard pattern.

According to a seventh aspect of the invention, in the invention according to the first aspect, the matching processing in the matching processing section is
This is word spotting processing.

[0016]

According to the first aspect of the invention, the acoustic feature amount of the acoustic signal input by the acoustic analysis unit is obtained for each frame, and the degree of the voice-likeness of the acoustic signal is obtained by the voice detection unit. At this time, if there is a section in the acoustic signal that is determined to have a low degree of voice likeness, the acoustic feature amount is converted to white noise by the noise conversion processing unit. That is, in the acoustic signal, a section with a low degree of soundness is a section with no sound or a low sound level. The fact that the acoustic feature amount is converted into white noise means that the spectrum expressing the acoustic feature amount is smoothed. Therefore, when there is no sound or the sound level is low, the spectrum as the acoustic feature amount is smoothed, and the influence of ambient noise is removed. In the matching processing unit, the time series of the feature amount that has undergone such processing is compared with the time series of the standard pattern stored in the standard pattern storage unit, and matching is performed. As a result, accurate voice recognition is performed regardless of the presence or absence of ambient noise.

According to the second aspect of the present invention, the cepstrum analysis using the cepstrum coefficient as the acoustic feature quantity is selected as the acoustic analysis method, and the white noise of the acoustic feature quantity is realized by reducing the cepstrum coefficient. . In the invention according to claim 3, the mel-cepstral analysis using the mel-cepstrum coefficient as the acoustic feature quantity is selected as the method of the acoustic analysis, and the mel-cepstral coefficient is reduced to realize white noise conversion of the acoustic feature quantity. ing. Further, in the invention according to claim 4, a spectrum analysis using a spectrum from which the spectrum inclination is removed is selected as the acoustic analysis method.
By reducing this spectrum, white noise of acoustic features is realized. Therefore, according to the second, third and fourth aspects of the present invention, accurate voice recognition is performed based on the stable acoustic feature amount, and the acoustic feature amount can be easily converted into white noise.

According to the fifth aspect of the present invention, when the degree of voice likeness in the voice detector is judged, the smaller the power of the input acoustic signal is, the lower the degree of voice likeness is. The degree of white noise conversion of the acoustic feature amount is determined according to the degree of power of the generated acoustic signal. That is, the smaller the power of the acoustic signal, the stronger the white noise becomes. As a result, more accurate voice recognition is performed.

In the invention according to claim 6, the standard pattern stored in the standard pattern storage section is generated through processing equivalent to white noise conversion of the acoustic feature quantity in the noise conversion processing section. Is easy to generate. Then, it is possible to prepare a standard pattern that is extremely close to the characteristic amount of the actually input acoustic signal, and more highly accurate voice recognition is performed.

In the seventh aspect of the invention, the matching section performs word spotting processing when matching the time series of the voice feature quantity and the time series of the standard pattern. As a result, when a certain word including a word or a syllable generated as a standard pattern is uttered, the word or the like included in the word is extracted and recognized.

[0021]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT An embodiment of the present invention will be described with reference to FIG. FIG. 1 is a block diagram of each unit. A voice input unit 1 for inputting a voice is provided with an A / D conversion unit 2 (A / D is ana
abbreviation of logto disital) is connected and this A / D
An acoustic preprocessing unit 3 and an acoustic analysis unit 4 are sequentially connected to the conversion unit 2. A voice detector 5 is also connected to the acoustic preprocessor 3, and the voice detector 5 and the acoustic analyzer 4 are connected.
The noise reduction processing unit 6 is connected to and. A standard pattern storage unit 7 is provided, the standard pattern storage unit 7 and the noise reduction processing unit 6 are connected to a matching processing unit 8, and the matching processing unit 8 is connected to a recognition result output unit 9. .

Here, the voice input unit 1 is, for example, a microphone, and has a structure in which an acoustic signal which is an analog signal input from the voice input unit 1 is output to the A / D conversion unit 2.

The A / D conversion unit 2 includes the voice input unit 1
It has a structure of converting an acoustic signal from the digital signal into a digital signal, and performing sampling, quantization and encoding. This A / D
The conditions for digital conversion in the conversion unit 2 are, for example, a sampling frequency of 16 kHz and a quantization bit number of 16 bits. At this time, sampling is performed after band limitation in the low frequency range. This is based on the sampling theorem and is sampled according to aliasing distortion.
) Is to prevent the occurrence of.

Next, the acoustic pre-processing unit 3 has a structure for emphasizing (pre-emphasis) the input acoustic signal converted into a digital signal by the A / D conversion unit 2. The acoustic preprocessing unit 3, H (z) = 1-order and a digital filter having a 1-Z transfer function to ¹ ......... formula 1, is constituted by the difference arithmetic circuit, and the like.

Next, the acoustic analysis unit 4 is a computing unit for extracting the feature amount of the input acoustic signal, and has a structure for extracting the spectral envelope of the short-time spectrum of the acoustic signal. In the present embodiment, a structure is used in which an LPC spectrum analysis is performed on an acoustic signal and a cepstrum vector (not including the 0th order) ct as a cepstrum coefficient is obtained for each frame. The acoustic signal analysis conditions in the acoustic analysis unit 4 are: frame period: 10 ms window period: 16 ms window function: Hamming window LPC analysis order: 14th order cepstrum order: 14th order.

Next, the voice detection unit 5 has a structure for obtaining the degree of voice-likeness of the acoustic signal based on the average power in each frame in the input acoustic signal which is high-frequency emphasized by the acoustic preprocessing unit 3. Is. The frame average power can be obtained from the zero-order autocorrelation coefficient during the LPC analysis. Here, in the voice detection unit 5, the relationship between the frame average power p and the degree of voice-likeness v obtained by the voice detection unit 5 is defined by the following Expression 2.

[0027]

[Equation 2]

P ₀ in the equation (2) is an experimentally determined constant, and a value slightly larger than the power value at the beginning and end of the voice section is used. As is clear from this formula,
The degree of soundness v is 0 ≦ v ≦ 1, the degree of soundness v is 1 when the frame average power p is sufficiently large, and the degree of soundness v is 0 when the frame average power p is 0. , Meanwhile, mean frame power p
As v becomes smaller, the degree of soundness v becomes monotonically lower.

Next, the noise conversion processing unit 6 has a structure for converting the acoustic feature amount obtained by the acoustic analysis unit 4 into white noise according to the degree of voice likeness obtained by the voice detection unit 5. is there. In the noise processing unit 6, the calculation process of c * t = vct ... Here, as described above, ct
Is a sound feature amount obtained by the sound analysis unit 4, that is, a cepstrum vector, and v is a degree of soundness obtained by the sound detection unit 5. And
c * t is a cepstrum vector that has been converted into white noise according to the degree of voice-likeness in the input acoustic signal. As is clear from this equation, the noise processing unit 6 determines the cepstrum vector c * t by the product of the cepstrum vector ct and the degree v of voice-likeness. Here, the cepstrum vector of white noise is 0, that is, 0.
Is a vector. Therefore, the lower the degree of voice-likeness v, the stronger the white noise of the cepstrum vector ct.

Next, in the standard pattern storage unit 7,
Many standard patterns such as words and syllables for executing voice recognition are stored. These standard patterns are input to the voice input unit 1, converted into the cepstrum vector ct by the acoustic analysis unit 4, and have the same contents as the time series of the cepstrum vector c * t subjected to the predetermined processing by the noise processing unit 4. Have
It is a cepstrum vector c * r generated through the same processing as the time series of the cepstrum vector c * t.

Next, the matching processing unit 8 stores the standard pattern stored in the standard pattern storage unit 7, that is, the time series of the cepstrum vector c * r, and the acoustic feature quantity that has been processed by the noise processing unit 6. That is, the structure is such that the matching process is executed with the time series of the cepstrum vector c * t. The matching process in the matching processing unit 8 is a matching process using the continuous DP method. At this time, as the distance measure, a distance measure such as a group delay spectrum distance measure is used.

Next, the recognition result output unit 9 has a structure for outputting the recognition result of the matching processing unit 8.
For example, the structure is such that the presence or absence of the corresponding word is output as a signal or display.

In such a configuration, the acoustic signal input to the voice input section 1 is digitally converted by the A / D conversion section 2, and is sampled, quantized and encoded. Then, the acoustic preprocessing unit 3 applies high-frequency emphasis to flatten the spectrum inclination. As a result, the dynamic range of the acoustic signal is compressed and the effective SNR (signal-to-quantization no
ise ratio: signal to quantization noise ratio) is increased.

Next, the high-frequency-emphasized acoustic signal is subjected to LPC cepstrum analysis by the acoustic analysis unit 4 and its characteristic amount is extracted as a cepstrum vector ct. At the same time, the voice detection unit 5 obtains the degree of voice-likeness v of the high-frequency-emphasized acoustic signal based on the average power p of each frame by the expression 2.

The cepstrum vector ct and the voice-likeness degree v thus obtained are calculated by the noise processing unit 6
And the white noise-ized cepstrum vector c * t is obtained by the calculation processing of the equation 3 in the noise-ized processing unit 6. Here, the cepstrum vector c * t processed by the noise conversion processing unit 6 is more strongly converted into white noise as the degree of voice likeness obtained by the voice detection unit 5 is lower. In other words, the low degree of voice-likeness means
Since it means that the section is silent or the voice level is low, the silent section or the section with a low voice level is converted into white noise, and the spectrum of the section is flattened.

Next, in the matching processing unit 8, the time series of the cepstrum vector c * t which has been processed by the noise processing unit 6 and the cepstrum vector c * r which is the standard pattern stored in the standard pattern storage unit 7. Are subjected to matching processing. The matching process at this time is performed by the continuous DP method which is word spotting. Therefore, the end-free voice recognition of the voice is performed. Then, the cepstrum vector c * t to be matched
And the time series of the cepstrum vector c * r are both white noise in a section with a low degree of voice-likeness, that is, a section with no sound or a low voice level, and the spectrum of that section is smoothed. ing. Therefore, the voice recognition is performed without being affected by the ambient noise, and the accuracy of the voice recognition is improved. Therefore, the matching processing unit 8
A high-accuracy recognition result is output from the recognition result output unit 9 that outputs the processing result of.

Here, a modification of the acoustic analysis unit 4 will be described. In the present embodiment, the acoustic analysis unit 4 that executes the LPC cepstrum analysis is provided as a method for obtaining the characteristic amount of the input acoustic signal, but the method for obtaining the characteristic amount of the acoustic signal is not limited to this. The acoustic analysis unit may use a technique such as analysis, mel cepstrum analysis, or spectrum analysis with spectrum tilt correction.
In short, any type of structure may be used as long as the feature quantity that can facilitate the white noise conversion process in the noise conversion processing unit 6 can be obtained. More specifically, in the mel cepstrum vector as the mel cepstrum coefficient, the 0 vector represents white noise, like the cepstrum vector. In addition, the spectrum analysis with the spectrum tilt correction is performed by performing spectrum tilt correction such as logarithmic conversion or correction for reducing the least-squares approximation straight line (exponential conversion) with respect to the spectrum obtained by the FFT or the bandpass filter bank. To be executed. As a result, in the corrected spectrum vector, the 0 vector represents white noise, like the cepstrum vector. Therefore, the standard pattern stored in the standard pattern storage unit can be applied to the apparatus of this embodiment as it is, only by the standard pattern subjected to the mel cepstrum analysis or the standard pattern subjected to the spectrum analysis with the spectrum tilt correction. .

Next, a modification of the voice detector 5 will be described. First, the voice detection unit 5 obtains the average power P of each frame as basic data for obtaining the degree of voice likeness in the acoustic signal.
Is obtained by LPC analysis, the voice detection unit 5 may be configured by partially sharing the structure of the acoustic analysis unit 4 that performs LPC cepstrum analysis. As another modification of the voice detection unit 5, the voice power is not used as the basic data for determining the degree of voice-likeness, but the number of zero crossings, the pitch frequency, the sharpness of the formant, the distance from each phoneme pattern, etc. are used. Is also good.

Next, a modification of the matching processing section 8 will be described. In the present embodiment, the matching processing unit 8 is configured to execute the continuous DP method, but the matching processing unit 8 may be configured to execute matching by another method using a state transition model or the like, and is not limited to such word spotting. Instead, the structure may be such that the isolated word voice recognition is executed.

[0040]

According to the first aspect of the present invention, an acoustic analysis unit that obtains an acoustic feature amount of an input acoustic signal by performing an acoustic analysis for each frame on the input acoustic signal, and an input acoustic signal On the other hand, a voice detection unit that obtains the degree of voice-likeness for each frame, a noise processing unit that whitens the acoustic feature quantity of the section that is determined to have a low degree of voice-likeness by this voice detection unit, and a voice standard. A standard pattern storage unit for storing patterns, and a matching processing unit for matching the time series of standard patterns stored in the standard pattern storage unit with the time series of acoustic feature values processed by the noise processing unit are provided. Therefore, if there is a section in the input acoustic signal that is determined to have a low degree of voice, that is, if there is a section with no sound or a low voice level, the acoustic feature amount is set to noise. The processing unit smoothes the spectrum that represents the acoustic features by converting it to white noise, thereby eliminating the effect of ambient noise on that section, and erroneously recognizing that the ambient noise features are mixed with the acoustic features. Therefore, there is an effect that the accuracy of voice recognition can be improved.

According to a second aspect of the present invention, in the first aspect of the invention, the acoustic analysis unit performs a cepstrum analysis for obtaining a cepstrum coefficient as an acoustic feature amount for each frame, and the noise conversion processing unit obtains the cepstrum coefficient. The acoustic feature quantity is converted into white noise by setting a small cepstrum coefficient of a frame having a low degree of speech-likeness. In the invention according to claim 3, in the invention according to claim 1, the acoustic analysis unit uses the melody as the acoustic feature quantity. The mel-cepstral analysis is performed to obtain the cepstrum coefficient for each frame, and the noise reduction processing unit converts the acoustic feature amount into white noise by setting the mel-cepstral coefficient of the frame with a low degree of speech likeness obtained by the voice detection unit to a small value, According to a fourth aspect of the present invention, in the first aspect of the present invention, the acoustic analysis unit uses a spectral gradient as an acoustic feature amount. The short-time spectrum correction is performed to remove performs a spectrum analysis for determining for each frame,
Since the noise conversion processing unit is configured to convert the acoustic feature quantity into white noise by setting the short-time spectrum of the frame with a low degree of soundness determined by the speech detection section to a small value, it is based on a stable acoustic feature quantity. Accurate voice recognition can be performed, therefore, the recognition accuracy of voice can be further improved, and the acoustic feature quantity can be easily converted into white noise. Therefore, arithmetic processing for converting to white noise can be performed. It is possible to achieve simplification.

According to a fifth aspect of the present invention, in the first aspect of the invention, the voice detection unit determines that the smaller the power of the input acoustic signal is, the lower the degree of voice-likeness is, and the noise reduction processing unit determines that voice. Since the detection unit is configured to strongly whiten the acoustic features as the degree of voice-likeness is determined to be low, it is easy and accurate to use the power of the acoustic signal as a parameter for determining the degree of voice-likeness. It is possible to realize the determination of the degree of soundness, and therefore, it is possible to simplify the arithmetic processing. Further, the smaller the power of an acoustic signal is, the more it is converted into white noise. It has an effect that it can contribute to the realization of recognition.

According to a sixth aspect of the present invention, in the first aspect of the invention, the standard pattern stored in the standard pattern storage section is subjected to processing equivalent to white noise conversion of the acoustic feature quantity in the noise processing section. Since it is the generated standard pattern, it is easy to generate the standard pattern, and it is possible to prepare a standard pattern that is extremely close to the characteristic amount of the acoustic signal actually input, and therefore a more accurate voice It has an effect that it can contribute to the realization of recognition.

According to a seventh aspect of the invention, in the invention according to the first aspect, the matching processing in the matching processing section is:
Since it is word spotting processing, when a certain word containing words or syllables generated as a standard pattern is uttered, it is possible to extract and recognize that word or the like contained in that word. This has the effect of preventing erroneous recognition results from occurring when there is no input voice.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention.

[Explanation of symbols]

 4 Acoustic Analysis Section 5 Voice Detection Section 6 Noise Reduction Processing Section 7 Standard Pattern Storage Section 8 Matching Processing Section

Claims (7)

[Claims] 1. An acoustic analysis unit that obtains an acoustic feature amount of an input acoustic signal by performing an acoustic analysis for each frame on the input acoustic signal, and an audio analysis unit for determining the soundness of the input acoustic signal for each frame. A voice detection unit that obtains a degree, a noise conversion processing unit that whitens the acoustic feature amount of a section determined to have a low degree of voice likeness by the voice detection unit, and a standard pattern storage unit that stores a standard pattern of voice. And a matching processing unit for matching the time series of the standard pattern stored in the standard pattern storage unit and the time series of the acoustic feature amount processed by the noise reduction processing unit. apparatus. 2. The acoustic analysis unit performs a cepstrum analysis for obtaining a cepstrum coefficient as an acoustic feature amount for each frame, and the noise reduction processing unit sets a small cepstrum coefficient for a frame having a low degree of voice likeness obtained by the voice detection unit. The speech recognition apparatus according to claim 1, wherein the acoustic feature amount is converted into white noise by performing the above. 3. The sound analysis unit performs a mel-cepstral analysis for obtaining a mel-cepstrum coefficient as an acoustic feature amount for each frame, and the noise-reduction processing unit includes a mel-cepstral coefficient of a frame with a low degree of soundness calculated by the sound detection unit. 2. The voice recognition apparatus according to claim 1, wherein the acoustic feature quantity is converted into white noise by setting a small value. 4. The sound analysis unit performs a spectrum analysis for obtaining a short-time spectrum for which correction is performed to remove a spectrum inclination as an acoustic feature amount for each frame, and the noise conversion processing unit performs speech likelihood determined by a voice detection unit. The speech recognition apparatus according to claim 1, wherein the acoustic feature quantity is converted into white noise by setting a short time spectrum of a frame having a low degree of noise to a small value. 5. The voice detection unit determines that the lower the power of the input acoustic signal is, the lower the likelihood of voice is, and the noise reduction processing unit determines that the voice detection unit is lower in the likelihood of voice. The speech recognition apparatus according to claim 1, wherein the acoustic feature quantity is strongly converted into white noise. 6. The standard pattern stored in the standard pattern storage unit is a standard pattern generated through a process equivalent to white noise conversion of the acoustic feature quantity in the noise conversion processing unit. 1. The voice recognition device according to 1. 7. The speech recognition apparatus according to claim 1, wherein the matching processing in the matching processing section is word spotting processing.