JP2006072127A - Voice recognition device and voice recognition method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To obtain a voice recognition device and a voice recognition method that does not cause significant increase in the hardware resources and having features of little instability in real environment and providing high recognition performance under various environments. <P>SOLUTION: A subtraction processing section is constituted of a signal storage section 8 which stores observed signals Y(f, m-1) to Y(f, m-L) equivalent to previous L frames; and a subtracting section 10, in which the signals that are obtained by respectively multiplying corresponding coefficients α<SB>1</SB>to α<SB>L</SB>of a subtracting coefficient storage section 9 to the observed signals Y(f, m-1) to Y(f, m-L) of the previous time frames stored in the signal storage section 8, and the resultant signals are subjected to power spectrum subtraction from the observed signals Y(f, m) of the current time frame; and the result is outputted as estimated signals Sset(f, m). A voice feature quantity extracting section 4 extracts a voice feature quantity from the inferred signals Sset(f, m). <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a speech recognition apparatus and a speech recognition method for recognizing speech input by a user in a reverberant environment such as a living room.

  Speech recognition technology plays an important role in realizing an excellent human interface. As a voice recognition apparatus to which the voice recognition technology is applied, an apparatus having a configuration as shown in FIG. 7 is conventionally provided (for example, Patent Document 1).

  The speech recognition apparatus includes a speech input unit 1 composed of a microphone that inputs speech, an A / D conversion unit 2 that performs A / D conversion on an output signal from the speech input unit 1, and an output from the A / D conversion unit 2. A frequency conversion unit 3 that converts the frequency of the signal, a voice feature amount extraction unit 4 that extracts a voice feature amount from an output signal from the frequency conversion unit 3, and a standard pattern for speech recognition created from the standard speech The standard pattern storage unit 7 and the speech feature amount of the input speech output from the speech feature amount extraction unit 4 are calculated with the similarity between the standard pattern stored in the standard pattern storage unit 7 and the recognition result 6 is obtained. The standard pattern to be stored in the standard pattern storage unit 7 collects standard feature patterns in advance and uses a modeling technique such as a hidden Markov model. Thing that was created is used.

  By the way, when the environment in which the device is used and the environment in which the standard pattern is created are different, there is a problem that the recognition rate decreases due to a difference between the user's voice and the standard pattern. In the speech recognition apparatus disclosed in No. 1, a plurality of standard patterns corresponding to the reverberation time of the environment are stored in the standard pattern storage unit 7, and the standard patterns are selected and used according to the environment. .

There is also provided a speech recognition device that recognizes an input speech by analyzing the input speech, estimating an inverse filter of the environment, and converting the input speech to a state not affected by the environment (for example, Non-Patent Document 1). )
JP 2004-117724 A (FIG. 1, paragraph number 0025) "Speech quality evaluation by clarity and recognition rate of dereverberation method using harmonic structure", Proc. Of the Acoustical Society of Japan, 611-612, published in March 2004

  In a method using a plurality of standard patterns, such as the speech recognition apparatus disclosed in Patent Document 1, there is a problem that the memory capacity increases to hold a plurality of standard patterns. In the speech recognition apparatus, since the ratio of the standard pattern to the entire memory capacity is large, there is a problem that the total memory capacity is doubled when a plurality of standard patterns are prepared.

  Further, the method of estimating the inverse filter of the environment from the input speech disclosed in Non-Patent Document 1 requires a large amount of input speech in order to estimate the inverse filter with sufficient accuracy. However, because the environmental characteristics change due to changes in the user's position and indoor temperature, it is difficult to obtain the required amount of input audio data within a certain period of time, resulting in an unstable inverse filter. Is learned and it is difficult to obtain sufficient recognition performance.

  The present invention has been made in view of the above-mentioned points, and the object of the present invention is not to involve a significant increase in hardware resources, and in various environments having features of less instability in a real environment. An object of the present invention is to provide a speech recognition apparatus method and speech recognition method that can obtain high recognition performance.

  In order to achieve the above object, in the invention of the speech recognition apparatus according to claim 1, a speech input unit that captures speech in a reverberant environment, and A / D conversion that performs A / D conversion on an output signal from the speech input unit A frequency conversion unit that frequency-converts an output signal from the A / D conversion unit and outputs an observation signal of a current time frame; stores an observation signal of a past predetermined time frame; From the observation signal, a subtraction processing unit that subtracts a signal obtained by multiplying the observation signal of the past predetermined time frame by a predetermined subtraction coefficient and outputs it as an estimation signal, and an estimation signal output from the subtraction processing unit Extracted from a voice feature quantity extraction unit that extracts voice feature quantities, a standard pattern storage unit that stores standard patterns created from standard voices in environments with different reverberation times, and the voice feature quantity extraction unit Features Characterized in that it includes a pattern matching unit for outputting a recognition result by determining the standard pattern and Noto similarity that is stored in the reference pattern memory.

  According to the speech recognition device of the first aspect of the present invention, it is not necessary to store a plurality of standard patterns, so that a large-capacity standard pattern storage unit is not required, and a delay is caused by reverberation from speech components captured in a reverberant environment. Since the speech feature amount can be extracted from the estimated signal having the same quality as that of the standard pattern by removing the speech component, it is possible to realize a speech recognition apparatus capable of performing speech recognition with high performance in various environments.

  In the speech recognition device according to claim 2, a speech input unit that captures speech in a reverberant environment, an A / D conversion unit that performs A / D conversion on an output signal from the speech input unit, and a signal from the A / D conversion unit A frequency converter that converts the frequency of the output signal and outputs an observation signal of the current time frame, and multiplies a predetermined signal corresponding to a past predetermined time frame by a predetermined subtraction coefficient from the observation signal of the current time frame. A subtraction processing unit that subtracts the obtained signal and outputs it as an estimation signal, and stores the estimation signal as a signal used for the predetermined signal, and a feature amount of speech from the estimation signal output from the subtraction processing unit A feature extraction unit, a standard pattern storage unit storing a standard pattern created from standard speech in an environment with different reverberation times, and a feature amount extracted from the speech feature extraction unit Said mark Characterized in that it includes a pattern matching unit for outputting a recognition result by determining the standard pattern and Noto similarity that is stored in the pattern storage unit.

  According to the speech recognition device of the second aspect of the present invention, it is not necessary to store a plurality of standard patterns, so that a large-capacity standard pattern storage unit is not required, and a delay is caused by reverberation from speech components captured in a reverberant environment. Speech features can be extracted from the estimated signal that has the same quality as the standard pattern by removing the speech component, so that speech recognition can be performed with high performance in various environments, especially up to past predetermined frames used for subtraction It is possible to realize a speech recognition apparatus in which a signal corresponding to the time frame is more accurate.

  According to a third aspect of the present invention, in the first or second aspect of the present invention, the subtraction coefficient is predetermined to a power ratio between the observation signal of the first frame and the observation signal of the predetermined frame of the transfer function of the reverberation environment. It is a value obtained by multiplying the coefficient of.

  According to the speech recognition device of the third aspect, the reduction coefficient can be derived more accurately, and as a result, the performance of speech recognition can be further improved.

  According to a fourth aspect of the present invention, there is provided the speech recognition apparatus according to any one of the first to third aspects, wherein the predetermined frame in the past used for subtraction is a plurality of frames, and a signal of each predetermined time frame is obtained. Multiplying a predetermined subtraction coefficient corresponding to each predetermined time frame, adding the multiplied results, and subtracting the addition result from the observation signal of the current time frame.

  According to the invention of the speech recognition apparatus of claim 4, dereverberation can be more accurately suppressed even for delay components of speech arriving at various delay times, and as a result, speech recognition performance can be further improved. .

  According to a fifth aspect of the present invention, there is provided a voice recognition apparatus according to any one of the first to fourth aspects, wherein the subtraction processing unit has a function of mitigating fluctuations of reverberation by smoothing in the subtraction processing.

  According to the speech recognition device of the fifth aspect, fluctuation of reverberation can be reduced, and as a result, speech recognition performance can be improved.

  According to a sixth aspect of the present invention, in the first to third aspects of the invention, the subtraction processing unit subtracts when one of the past predetermined time frames is used for the output of the estimated signal. The past predetermined time frame is selected for each frequency band or frequency for processing.

  According to the invention of the speech recognition apparatus of claim 6, dereverberation can be more accurately suppressed even for delay components having different delay times depending on frequencies, and as a result, the performance of speech recognition can be improved.

  The speech recognition method according to claim 7 is a speech recognition method for obtaining a recognition result by obtaining a similarity between a feature amount of speech captured in a reverberant environment and a standard pattern created from the standard speech, Subtract the signal obtained by multiplying the observation signal of the past predetermined time frame by the predetermined subtraction coefficient from the observation signal of the current time frame obtained by frequency conversion of the audio signal captured below and the frequency conversion And obtaining a presumed signal, extracting a feature amount of speech from the presumed signal, and performing speech recognition from a similarity between the extracted feature amount and the standard pattern. .

  According to the speech recognition method of the present invention, the speech feature amount is extracted from the estimated signal having the same quality as the standard pattern by removing the speech component delayed by the reverberation from the speech component captured in the reverberant environment. Therefore, it is possible to improve the performance of voice recognition in various environments without significantly increasing hardware.

  The speech recognition method according to claim 8 is a speech recognition method for obtaining a recognition result by obtaining a similarity between a feature amount of speech captured in a reverberant environment and a standard pattern created from the standard speech, Obtained by multiplying a predetermined signal corresponding to a past predetermined time frame by a predetermined subtraction coefficient from the process of frequency converting the audio signal captured below and the observation signal of the current time frame obtained by frequency conversion. A process of subtracting signals to obtain an estimated signal and storing the estimated signal as a signal to be used for the predetermined signal; a process of extracting a speech feature quantity from the estimated signal; and the extracted feature quantity and the standard And a process of performing speech recognition based on the similarity to the pattern.

  According to the speech recognition method of the present invention, the speech feature amount is extracted from the estimated signal having the same quality as the standard pattern by removing the speech component delayed by the reverberation from the speech component captured in the reverberant environment. Therefore, it is possible to improve speech recognition performance in various environments without significantly increasing hardware, and in particular, signals corresponding to time frames up to a predetermined frame used for subtraction become more accurate. .

  The invention of the speech recognition apparatus eliminates the need for storing a number of standard patterns, eliminates the need for a large-capacity standard pattern storage unit, and eliminates speech components delayed by reverberation from speech components captured in a reverberant environment. Thus, it is possible to extract a voice feature amount from an estimated signal having the same quality as that of the standard pattern, so that it is possible to realize a voice recognition device that can perform voice recognition with high performance in various environments.

  In addition, the invention of the speech recognition method can extract a speech feature amount from an estimated signal having the same quality as a standard pattern by removing a speech component delayed by reverberation from a speech component captured in a reverberant environment. The voice recognition performance can be improved in various environments without a significant increase in wear.

  Embodiments of the present invention will be described below.

(Embodiment 1)
FIG. 1 shows the configuration of this embodiment. In this embodiment, the observed signal Y (f, m) frequency-converted by the frequency converter 3 is a past time frame up to a predetermined frame, for example, L frames, That is, the signal storage unit 8 that stores Y (f, m−1) to Y (f, m−L) and the L past observation signals Y (f, m−) stored in the signal storage unit 8. 1) to Y (f, m−L) subtraction coefficient storage unit 9 storing _L subtraction coefficients α _{1 to} α _L corresponding to each of them, and each past observation stored in the signal storage unit 8 The observation signal observation signals Y (f, m−1) to Y (f, m) stored in the subtraction coefficient storage unit 9 are converted into signals Y (f, m−1) to Y (f, m−L). multiplied by the subtraction factor alpha ₁ to? _L corresponding to -L), a signal obtained by multiplying the current time frame which is outputted from the frequency converter 3 A subtraction unit 10 that subtracts the power spectrum by subtracting from the observed signal of the audio signal, and the subtraction result output from the subtraction unit 10 is used as an estimation signal for extracting a voice feature amount. It is characterized in that it is output to the unit 4.

  The signal storage unit 8 stores the observation signals output from the frequency conversion unit 3 for L frames as described above. Each time an observation signal is input, the observation signal of the oldest frame is deleted, and a new frame is deleted. The observation signal is memorized.

  The voice input unit 1 composed of a microphone captures voice in a reverberant environment as in the conventional example, and the voice signal output from the voice input unit 1 is A / D converted by the A / D converter 2 and then the frequency converter. 3 is frequency-converted.

  The pattern matching unit 5 obtains the similarity between the voice feature amount extracted by the voice feature amount extraction unit 4 and the standard pattern of the standard voice stored in the standard pattern storage unit 7, and the recognition result corresponding to the similarity degree. 6 is output. The standard pattern storage unit 7 does not store a standard pattern under a plurality of environments, but only stores a voice pattern based on one standard voice, and stores a plurality of standard patterns as in the conventional example. Compared to a memory having a significantly smaller storage capacity.

  Next, the operation of this embodiment will be described.

  First, the user measures a transfer function from the user position in the environment where the apparatus is used to the voice input unit 1 as a microphone. FIG. 2 shows an example of the result of the measurement of the transfer function h (t).

Next, for example, a value obtained by multiplying the power ratio between frames as shown in Expression (1) by a predetermined constant β is α _p . Here, the window width W is set to the same value as the window length in the frequency conversion section 3, for example, a window shift width (width in FIG. 2 from t = 0 to t = T ₁₎ is likewise in the frequency converter 3 The same value as the window shift width.

FIG. 3 shows an example of the relationship between the obtained α _p value and each frame.

The α _p calculated as described above is stored in the subtraction coefficient storage unit 9 in advance. Since an appropriate means may be used as the means for storing, it is omitted here.

The α _p calculated as described above is stored in the subtraction coefficient storage unit 9 for L frames (α _{1 to} α _L ), and the preparation is completed.

Now, the operation of the speech recognition apparatus of this embodiment for recognizing input speech is as follows. In the state where the past observation signals Y (f, m−1) to Y (f, m−L) for L frames are already stored in the signal storage unit 8, the observation signal Y corresponding to the current time frame is already stored. When (f, m) is input to the subtraction unit 10, the subtraction unit 10 sequentially reads the observation signals Y (f, m−1) to Y (f, m−L) from the signal storage unit 8 and stores the subtraction coefficient. The subtracting coefficients α _{1 to} α _L for L frames are read from the unit 9, for example, subtraction processing is performed in the power spectrum region as shown in Expression (2-1), and the estimated signal Sest (f, m) is output.

  If the result of subtraction (shown by equation (2-2)) becomes a negative value, flooring processing as in equation (3) or zero signal as in equation (4) is performed. May be.

Set (f, m) = Y (f, m) × 0.5 (3)
Set (f, m) = Y (f, m) × 0 (4)
As described above, the speech recognition apparatus according to the present embodiment includes the subtraction processing unit including the signal storage unit 8, the subtraction coefficient storage unit 9, and the subtraction unit 10, so that the voice input is delayed by being reflected on the wall or the floor. The speech component input to the unit 1 can be removed from the observation signal, so that the estimated signal output from the subtracting unit 10 by the subtraction process becomes the same quality as the standard pattern, and the speech feature amount extracting unit from this estimated signal The pattern matching unit 5 calculates the similarity between the feature amount extracted in 4 and the standard pattern based on the standard voice, and the recognition result 6 output based on the result is obtained with a high recognition rate. Will improve.

  In addition, since the subtraction process is performed by a plurality of frames, it is possible to cope with a plurality of delay components due to various reflection paths such as a floor and a wall, and the recognition performance can be further improved.

Furthermore, since the subtraction coefficient α _p is calculated from the transfer function, more accurate subtraction processing can be performed, resulting in improved recognition performance.

Note that α _p is calculated as described above based on the transfer function measured in an environment where the speech recognition apparatus is actually used. For example, α _p calculated in a plurality of environments where the speech recognition apparatus can be used. It is good also as an average value.

Further, if the estimated signal obtained by the subtraction process is subjected to inverse frequency conversion, an audio signal with reduced reverberation can be obtained, so that it can be applied to hands-free telephones, intercoms, etc. in addition to the voice recognition device.
(Embodiment 2)
In the present embodiment, in addition to the configuration of the first embodiment, as shown in FIG. 4, past frame observation signals Y (f, m) stored in the signal storage unit 8 between the signal storage unit 8 and the subtraction unit 10. -P) is characterized in that a filter unit 11 is provided for smoothing in the frequency axis direction.

  Since the other configuration is the same as that of the first embodiment, the same components as those of the first embodiment are denoted by the same reference numerals, and the description of the common components is omitted.

  Thus, when the subtraction unit 10 reads the observation signal Y (f, mp) of the past frame stored in the signal storage unit 8, the signal Y (f-1, m−p) and Y (f + 1, m−p) are read from the signal storage unit 8, and for example, a signal subjected to smoothing processing as shown in Expression (5) is output to the subtraction unit 10.

  FIG. 5 shows the state of the frequency when the audio signal travels through the space. FIG. 5A shows the case where there is no frequency fluctuation, and FIG. 5B shows the case where the audio is reflected by a wall or the like. A case where the frequency fluctuates is shown.

As shown in FIG. 5 (a) (i), for example, K frames before the speech user encounters a frequency f ₀ of the elements (● mark) is a delay of K frames through the wall reflections, etc. transmitted spatial When the current frame is attenuated and mixed, an element (▲ mark) of the frequency f ₀ of the voice generated by the user is observed in the current frame as shown in (b) of FIG. Therefore, the observation signal becomes a sound in which both are mixed, and the sound with reverberation as described above is observed.

On the other hand, when the sound is reflected by a wall or the like, as shown in (b) of FIG. 5B, the element (● mark) of the frequency f ₀ of the sound generated by the user before the K frame is the fluctuation of the frequency. Thus, as shown in (b) of FIG. 5B, for example, it is mixed in the element (Δ mark) of the frequency f ₀ −1 of the current frame, but is based on the above-described equation (5) by the filter unit 11. Smoothing process.

  Note that the processing operation until the recognition result is obtained from the estimated signal Sest (f, m) obtained by the subtraction processing of the subtracting unit 10 is the same as that in the first embodiment, and thus the description thereof is omitted.

As described above, in the speech recognition apparatus according to the present embodiment, even when the delay component due to reverberation fluctuates in the frequency axis direction, the observation signal Y (f, mp) of the past time frame used for subtraction is filtered. 11, the delay component can be subtracted by the subtracting unit 10 from the frequency component of the current observation signal Y (f, m). Components can be removed, resulting in improved recognition performance.
(Embodiment 3)
In the first embodiment described above, the observation signal of the past time frame is used in the subtraction process, but this embodiment is characterized in that the past estimated signal is used for subtraction instead of the observation signal.

  That is, in the present embodiment, as shown in FIG. 6, the estimated signal Sest (f, m) output as the subtraction result in the subtraction unit 10 is output to the audio feature amount extraction unit 4 and also output to the signal storage unit 8. It is supposed to be.

  The signal storage unit 8 stores the estimated signal output from the subtracting unit 10 over the past L frames, and sequentially outputs it according to the reading from the subtracting unit 10.

  Since other configurations are the same as those of the first embodiment, common constituent elements are denoted by the same reference numerals, and description of the common constituent elements is omitted.

Thus, the subtraction unit 10 of the speech recognition apparatus according to the present embodiment uses the subtraction coefficient α _p read from the subtraction coefficient storage unit 9 corresponding to the estimated signal Set (f, m−p) read from the signal storage unit 8. Subtraction processing similar to that in the first embodiment is performed.

  Thus, since the signal used for the subtraction processing in the subtraction unit 10 is not the observation signal Y (f, mp) but the estimated signal Sest (f, mp) from which the delay component due to reflection has been removed. In the speech recognition apparatus of the present embodiment, only the delay component can be subtracted more accurately, and as a result, the recognition performance is improved.

In the present embodiment, initially, the signal storage unit 8 is in a state of a zero signal in which the estimated signal Est (f, mp) corresponding to the past time frame is not stored, so the subtraction result at this time is (F, m) = Y (f, m) × e ^{jarg (Y (f, m))} , which is stored in the signal storage unit 8 as the estimated signal Sest (f, m−p) one frame before. Will be. Thereafter, until the estimated signals Sest (f, m-1) to Sest (f, m-L) for L frames are stored, for the past time frames in which the estimated signals are not stored, the zero signal is estimated for that frame. Used as a signal.

In the present embodiment, a filter unit 11 similar to that of the second embodiment may be provided in order to perform smoothing processing on the estimated signal Sest (f, mp) output from the signal storage unit 11.
(Embodiment 4)
First, in the first embodiment, when only one frame is used as the past observation signal used for the subtraction process, the subtraction processing unit uses the past observation signal Y (f, m−K) of the past K frame. To subtract. However, K is one of 1, 2,... L.

Here, in the present embodiment, the point where the above-described value K is changed for each frequency band or for each frequency is different from the first embodiment. In the present embodiment, for example, different K ₁ , K ₂ , and K 3 are selected as K.

For frequencies 0 to f ₁ , the observation signal Y (0, m−K ₁ ) before K ₁ frame,
Y (1, m−K ₁ ),..., Y (f ₁ , m−K ₁ ) is stored in the signal storage unit 8. For the frequency bands f1 + 1 to f2, the observation signals Y (f ₁ +1, m−K ₂ ), Y (f ₁ +2, m−K ₂ ),..., Y (f ₂ , m−K ₂ ) before the K2 frame. ) Is stored in the signal storage unit 8. For the frequency f ₂ +1, the observation signal Y (f ₂ +1, m−K ₃ ) before K ₃ frames is stored in the signal storage unit 8.

  Thus, when the signal storage unit 8 receives the frequency-converted signal from the frequency conversion unit 3, the signal storage unit 8 stores each observation signal as described above by delaying it by a predetermined number of time frames for each frequency band. When the readout signal is received from the unit 10, the stored observation signals are sequentially output to the subtraction unit 10. The attenuation coefficient stored in the attenuation coefficient storage unit 9 is also stored in a corresponding manner.

  Since other configurations and operations are the same as those of the first embodiment, the description of the configurations will be omitted with reference to FIG.

  As described above, in the speech recognition apparatus according to the present embodiment, a signal corresponding to a delay component due to voice reflection or the like can be stored with a delay amount changed for each frequency band. Can be subtracted with high accuracy, and as a result, the recognition performance is improved.

Note that the signal multiplied by the subtraction coefficient α _p is used instead of the observation signal Y (f, m−p), as in the third embodiment, using the estimated signal Sest (f−m−p) corresponding to a past predetermined frame. Also good.

FIG. 6 is a circuit configuration diagram of the first and fourth embodiments. It is a measurement example figure of a transfer function in a use environment. FIG. 6 is an explanatory diagram of a relationship between an attenuation coefficient value and a frame used in the first embodiment. 6 is a circuit configuration diagram of Embodiment 2. FIG. It is explanatory drawing of the fluctuation of the frequency at the time of an audio signal propagating through space. 6 is a circuit configuration diagram of Embodiment 3. FIG. It is a circuit block diagram of a prior art example.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Voice input part 2 A / D conversion part 3 Frequency conversion part 4 Voice feature-value extraction part 5 Pattern collation part 6 Recognition result 7 Standard pattern storage part 8 Signal storage part 8
9 Subtraction coefficient storage unit 9
10 Subtraction part

Claims (8)

An audio input unit that captures audio in a reverberant environment, an A / D converter that performs A / D conversion on the output signal from the audio input unit, and frequency-converts the output signal from the A / D converter to A frequency converter that outputs an observation signal of a time frame, stores an observation signal of a past predetermined time frame, and multiplies the observation signal of the past predetermined time frame by a predetermined subtraction coefficient from the observation signal of the current time frame. A subtraction processing unit that subtracts the signal obtained in this way and outputs it as an estimation signal; a speech feature extraction unit that extracts a speech feature amount from the estimation signal output from the subtraction processing unit; A standard pattern storage unit storing a standard pattern created from a standard voice in the system, a feature amount extracted from the voice feature amount extraction unit, and a standard pattern stored in the standard pattern storage unit Similar Speech recognition apparatus characterized by and a pattern matching unit for outputting a recognition result seeking. An audio input unit that captures audio in a reverberant environment, an A / D converter that performs A / D conversion on the output signal from the audio input unit, and frequency-converts the output signal from the A / D converter to Estimated by subtracting the signal obtained by multiplying the predetermined signal corresponding to the past predetermined time frame by the predetermined subtraction coefficient from the frequency converter that outputs the observation signal of the time frame and the observation signal of the current time frame A subtraction processing unit that outputs the signal as a signal and stores the estimation signal as a signal used for the predetermined signal; and a voice feature amount extraction unit that extracts a voice feature amount from the estimation signal output from the subtraction processing unit; A standard pattern storage unit storing a standard pattern created from standard speech in an environment with different reverberation times, a feature amount extracted from the speech feature amount extraction unit, and a standard pattern storage unit Speech recognition apparatus characterized by and a pattern matching unit for outputting a recognition result determining the standard pattern and Noto similarity. 3. The subtraction coefficient is a value obtained by multiplying a power ratio between an observation signal of a first frame and an observation signal of a predetermined frame of the transfer function of the reverberant environment by a predetermined coefficient. Voice recognition device. The past predetermined time frame used for subtraction is a plurality of frames, and a signal corresponding to each predetermined time frame is multiplied by a predetermined subtraction coefficient corresponding to each predetermined time frame and the result of multiplication is added. 4. The speech recognition apparatus according to claim 1, wherein the addition result is subtracted from the observation signal of the current time frame. The speech recognition apparatus according to claim 1, wherein the subtraction processing unit has a function of mitigating fluctuations of reverberation by smoothing in the subtraction process. The subtraction processing unit selects the past predetermined time frame for each frequency band or frequency for which the subtraction processing is performed when one past predetermined time frame is used for the output of the estimation signal. The speech recognition apparatus according to claim 1. A speech recognition method that obtains a recognition result by obtaining the similarity between a feature amount of speech captured in a reverberant environment and a standard pattern created from the standard speech, and a process of frequency-converting the speech signal captured in the reverberant environment And subtracting a signal obtained by multiplying an observation signal of a past predetermined time frame by a predetermined subtraction coefficient from an observation signal of a current time frame obtained by frequency conversion, and obtaining an estimation signal; A speech recognition method, comprising: a step of extracting a speech feature amount from a signal; and a step of performing speech recognition based on a similarity between the extracted feature amount and the standard pattern. A speech recognition method that obtains a recognition result by obtaining the similarity between a feature amount of speech captured in a reverberant environment and a standard pattern created from standard speech, and a process of frequency-converting a speech signal captured in a reverberant environment And subtracting a signal obtained by multiplying a predetermined signal corresponding to a past predetermined time frame by a predetermined subtraction coefficient from an observation signal of the current time frame obtained by frequency conversion, and obtaining an estimated signal; and Storing the estimated signal as a signal to be used for the predetermined signal; extracting a feature amount of speech from the estimated signal; and performing speech recognition from a similarity between the extracted feature amount and the standard pattern A speech recognition method comprising:

Images