JP2005070377A - Device and method for speech recognition, and speech recognition processing program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve a speech recognition rate more by performing recognition while seizing redundant variation of a frequency-base system feature parameter and momentary variation of a power-system feature parameter although feature vectors are in the same cycle by making the window length of the power-system feature parameter representing a feature of an abrupt variation part of a consonant most shorter than the window length of the frequency-base-system feature parameter representing a feature of a redundant part of a vowel most. <P>SOLUTION: The upper part shows a speech section of fixed length to be analyzed in an inputted speech waveform; and the lateral axis is a time base and the longitudinal axis represents the amplitude (energy) of the speech waveform. The intermediate stage shows four time windows F(i) (i=1 to 4) for the frequency-base-system feature parameter obtained by shifting at equal intervals of a shift length FS within a range of an analytic frame, and a time window P(i) for the power-system feature parameter. Those time windows have window lengths fL and pL of fixed lengths respectively, the window length pL of the time window P(i) for the power-system feature parameter being shorter than the window length fL of the time window F(i) for the frequency-axis-system feature parameter. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a voice recognition device, a voice recognition method, and a voice recognition processing program.

  In recent years, research on speech recognition for recognizing human speech by a machine has been actively conducted, and it has partially reached the practical range. As a speech recognition method, a method is adopted in which an input speech and a standard pattern given in advance are compared, and the one with the highest similarity is selected and output.

  This standard pattern is different depending on the speech recognition method employed. For example, in a speech recognition method called DTW (Dynamic Time Warpimg) using DP (Dynamical Programming), A time series of typical speech feature values is used as a standard pattern.

  In addition, the speech recognition method using HMM (Hidden Markov Model) uses a statistical method, and each word included in speech data is in a plurality of states with relatively few phonemes. In this case, a standard pattern is used in which a state transition probability for each word and a probability of outputting a feature amount input from each state are given as parameters. At present, this HMM is widely used as a central speech recognition method.

  In such a speech recognition method, information necessary for speech recognition, that is, a speech feature amount is extracted from a speech signal uttered by a human and compared with a standard pattern. The quality of performance affects the performance (processing speed and recognition rate) related to overall speech recognition.

  A conventional method for extracting speech feature values of speech input is set by shifting a time window having a constant window length, such as a hamming window, within a range of an analysis frame by shifting the time series by a certain time. A method is used in which speech samples are extracted from speech waveforms in analysis frames sequentially extracted by each time window, and speech feature values of the speech samples in the time window are extracted.

  Such a voice feature amount is a voice feature amount obtained by converting a predetermined number of voice samples sequentially cut out for each time window onto the frequency axis (hereinafter referred to as a frequency axis system feature parameter), Combining with the voice feature quantity (hereinafter referred to as power system feature parameter) obtained by taking the sum of squares of the voice amplitude quantized by linear PCM (Pulse Code Modulation; pulse coding) or its logarithm. Is obtained.

  For example, the difference between the frequency axis system characteristic parameter 12 component (12 dimensions), the power system characteristic parameter 1 component (1 dimension), and each component of the immediately preceding time window, that is, the frequency axis system characteristic parameter 12 component ( It is conceivable that the feature amount is composed of a total of 26 components, ie, 12 dimensions) and one power system feature parameter component (one dimension) as a 26-dimensional vector quantity.

  As described above, the voice feature amount of the input voice is extracted based on the voice samples cut out in each time window by setting the time windows so as to partially overlap within the range of the analysis frame. If the time window overlap is set to be short, the number of time windows decreases, so the number of audio samples cut out from the input speech decreases, the number of voice feature extraction processes, and the number of subsequent voice recognition processes. Are both kept low, and the processing speed can be increased. On the other hand, since the sampling becomes rough, the statistical property is deteriorated and the speech recognition rate is lowered.

On the other hand, if the time window overlaps is set longer, the number of time windows increases, so the sampling becomes finer and the number of audio samples cut out from the input speech increases, thus improving statistics and enabling accurate speech recognition. However, on the other hand, the number of voice feature extraction processes and the number of subsequent voice recognition processes increase, resulting in a decrease in processing speed.
That is, when the processing speed is increased, the recognition rate decreases, and when the recognition rate is increased, the processing speed decreases, and it is difficult to improve the processing speed and the recognition rate at the same time.

In order to solve this problem, the recognition processing speed and the recognition rate are set so that speech recognition can be performed most efficiently by differentiating the period for extracting the frequency axis system characteristic parameter and the period for extracting the power system characteristic parameter. Proposals for optimization have been made (see, for example, Patent Document 1).
JP 2000-356790 A

  Human sound recognition is known to be more sensitive to sudden sounds than to stationary sounds. In contrast to the stationary sound, the instantaneous sound can be clearly distinguished even with a slight difference. In other words, human beings recognize the content of pronunciation by sensitively and finely discriminating sudden sounds. To deal with this, the window length or the time window shift width can be simply made fine, but this leads to an increase in the amount of processing. In addition, if the time window is shortened, the power system characteristic parameter that best represents the characteristics of the sudden fluctuation part of the consonant part of the speech can be extracted, but the characteristic of the redundant part of the vowel part is best expressed. There is a fear that the characteristic parameter of the frequency axis system cannot be extracted.

  Accordingly, an object of the present invention is to make it possible to discriminate and recognize sudden sounds as well as humans without increasing the processing amount.

  In order to achieve the above object, the present invention sets a time window of a predetermined length with respect to an analysis target voice at a predetermined cycle, and uses the time window as a processing unit, a frequency axis system characteristic parameter related to the frequency of the voice, and a voice Extracting a feature quantity consisting of power system feature parameters related to the amplitude of the signal, and recognizing the analysis target speech based on the extracted feature quantity, the frequency of the length of the time window for extracting only the power system feature parameters It is characterized in that the feature quantity is extracted by making it shorter than the length of the time window for extracting only the axis system feature parameters.

The length of the time window for extracting only the frequency axis system characteristic parameter is preferably equal to or longer than the period of the basic pitch component in the vowel part of the input speech.
Moreover, it is desirable to perform speech recognition by calculating a difference feature parameter representing a difference between temporally adjacent feature parameters among feature parameters that are sequentially generated in a predetermined cycle.

  And it further has a band pass filter means for passing only the band of 3 kHz to 8 kHz with respect to the analysis target voice, and sequentially extracts the power system characteristic parameters of the input voice through the band pass filter means at the predetermined period, The extracted difference between the power system characteristic parameters adjacent in time may be used for the voice recognition means.

  With the above configuration, the present invention best represents the characteristics of the sudden fluctuation part of the consonant with respect to the window length of the frequency axis system characteristic parameter that best represents the characteristic of the redundant part of the vowel. By shortening the window length of the power system feature parameter, the frequency axis feature parameter is redundant while the power axis feature parameter is recognized while capturing the instantaneous change. This can be done and the speech recognition rate is further improved.

FIG. 1 is a block diagram showing an internal configuration of a speech recognition apparatus 1 using an HMM model according to an embodiment of the present invention.
As shown in FIG. 1, the speech recognition apparatus 1 includes a standard pattern (preliminarily stored in a time window position setting unit 11, a speech feature amount extraction unit 12, a comparison unit 13, and a storage device (not shown)). HMM models 141 to 14n).

FIG. 2 is a diagram illustrating how the time window position setting unit 11 sets a time window for an input speech waveform. As an example, four time windows are cut out from a speech section (analysis frame) to be analyzed. Is shown.
The upper part of FIG. 2 shows a fixed-length speech section that is an analysis target of the input speech waveform, the horizontal axis represents the time axis, and the vertical axis represents the amplitude (energy) of the speech waveform. The middle part of FIG. 2 shows four frequency axis system feature parameter time windows F (i) (i = 1 to 4) and power system feature parameters that are shifted at equal intervals by the shift length FS within the range of the analysis frame. A time window P (i) is shown. Each of these time windows has window lengths fL and pL having a certain length, and as is apparent from FIG. 2, the window length pL of the power system characteristic parameter time window P (i) is the frequency axis system characteristic parameter. It is shorter than the window length fL of the working time window F (i).

  Also, the frequency axis feature parameter time window F (i) is such that adjacent time windows partially overlap, and the power feature parameter time window P (i) is the same as the period FS in this embodiment. Further, the start position is delayed by an offset amount from the time axis F for the frequency axis system characteristic parameter. The offset amount is such that the start position of the power system characteristic parameter time window P (i) is always at the same relative time position with respect to the start position of the frequency axis system characteristic parameter time window F (i). It may be “0”.

In the time window position setting unit 11 shown in FIG. 1, the window lengths fL, pL, the offset amount, and the time window setting period FS shown in FIG. 2 are set.
The time window position setting unit 11 sequentially sets the time windows F (i) and P (i) according to the window lengths fL, pL, the amount of offset, and the time window setting period FS set within the range of the analysis frame. Then, a start control signal for starting the extraction of the voice feature quantity for the frequency axis system parameter is output to the voice feature quantity extraction unit 12.

Thus, after a time corresponding to the offset amount has elapsed, a start control signal for starting extraction of the speech feature amount for the power system parameter is output to the speech feature amount extraction unit 12. After this, after the window length pL, an end control signal for ending the extraction of the speech feature amount for the power system parameter is output to the speech feature amount extraction unit 12.
Then, after the extraction of the voice feature quantity for the frequency axis system parameter is started, after the window length fL, an end control signal for terminating the voice feature quantity extraction is output to the voice feature quantity extraction unit 12. This series of operations is repeated for every cycle FS until the analysis frame ends.

  Based on the start control signal and the end control signal of the time window F (i) input from the time window position setting unit 11, the voice feature amount extraction unit 12 converts the input voice into the time window F (i) in the analysis frame. The frequency axis system characteristic parameter f (i) (for example, a D-dimensional vector amount) is extracted from the audio data d (n) in the time window F (i), and the audio data in the time window P (i) is extracted. The power system characteristic parameter p (i) (one-dimensional vector quantity) is calculated from d (n).

  The power system characteristic parameter p (i) is a feature quantity related to the amplitude of the voice, and is a one-dimensional vector quantity obtained with a relatively small calculation quantity, for example, calculating the sum of squares of the voice data d (n) and its logarithm. is there. The frequency axis system characteristic parameter f (i) is, for example, a feature quantity related to the frequency of speech called cepstrum or mel cepstrum, and FT (Fourier Transform) or logarithmic transformation for speech data d (n). , A D-dimensional vector quantity composed of a plurality of (for example, D) calculation results such as mel-axis transformation.

  Further, as shown in FIG. 3, the audio feature quantity extraction unit 12 is a difference Δf (i) between frequency axis system feature parameters adjacent in time (vector quantity having the same dimension as f (i), for example, D dimension). Or the difference Δp (i) (one-dimensional vector quantity) between adjacent power system feature parameters is calculated, and these differences Δf (i) and Δp (i) are added to the unit voice feature quantity. In this embodiment, ΔΔp (i), which is a further difference of the power system feature parameter difference Δp (i), is calculated and added to the unit voice feature amount.

  Here, Δp (i) indicates a dynamic characteristic of the power system characteristic parameter, and with respect to a frequency axis system characteristic parameter time window F (i) that must include a basic pitch component that is a characteristic of a vowel, By having a unique dedicated window length pL with a short window width, the characteristics of the consonant part that changes sharply in a short time are well expressed, and a good recognition result can be obtained.

    The comparison unit 13 recognizes the input voice by comparing and comparing the unit voice feature amount composed of various parameters extracted in the frame and a standard pattern stored in advance in a storage device (not shown). And output the recognition result. In this embodiment, speech recognition is performed by a statistical method based on HMM.

  Here, a speech recognition method based on a statistical method using the HMM will be described. Voice recognition is performed based on a standard pattern called HMM models 141 to 14n given in advance using the feature quantity of the input voice extracted by the voice feature quantity extraction unit 12. The HMM models 141 to 14n represent each word included in the speech data in a plurality of states having relatively few phonemes, and output a state transition probability and a feature amount input from each state for each word. Probability is given as a parameter. The comparison unit 13 calculates the likelihood (probability) of which HMM model outputs the given speech feature amount with the highest probability among the HMM models 141 to 14n, and the HMM that maximizes the probability. The word corresponding to the model is output as a speech recognition result.

Next, the operation will be described. FIG. 4 is a flowchart for explaining the speech recognition processing of the speech recognition apparatus 1.
First, when speech is input to the speech recognition apparatus 1, the input speech is sampled by sampling at a predetermined sampling interval by an A / D converter (not shown) provided in the input stage. PCM-encoded based on the linear PCM method for performing conversion to speech data d (n), which is output to the time window position setting unit 11 and the speech feature amount extraction unit 12.

  When the audio data d (n) is input to the time window position setting unit 11, the time windows F (i) and P (i) such as a Hamming window are specified based on the window lengths fL, pL, and the amount of offset. The start control signal / end control signal is output from the time window position setting unit 11 to the audio feature amount extraction unit 12 (step S40). Next, the audio feature quantity extraction unit 12 extracts audio data d (n) in the time window F (i) based on the control signal output in step S40, and the start position of the time window F (i). P (i) is cut out from the position where the time corresponding to the offset amount has elapsed (step S41).

  The speech feature amount extraction unit 12 calculates the power system feature parameter p (i) by taking the square sum of the speech data d (n) or the logarithm thereof in the time window P (i) cut out in step S41. Further, a frequency axis system characteristic parameter f (i) obtained by converting the audio data d (n) in F (i) onto the frequency axis by FT or the like is extracted. (Step S42). Along with the extracted power system feature parameter p (i), the audio feature quantity of the time window F (i) is also stored (step S43).

  Next, it is determined whether or not the voice segment cut out in this time window F (i) has ended (step S44). If not completed, the process proceeds to step S45, the time windows F (i) and P (i) are shifted to the time windows F (i + 1) and P (i + 1), and the process returns to step S41.

  If it is determined in step S44 that the voice section has ended in the time window F (i), the voice feature values stored for each time window F (i) are arranged in time series in step S46.

  Next, the speech feature extraction unit 12 calculates a difference (Δ and ΔΔ) by the following formula to generate a feature vector sequence for the speech feature amounts arranged in time series in step S46 (θ is a frame before and after considering). number).

  Using the feature vector sequence (f (i), Δp (i), Δf (i), ΔΔp (i)) thus obtained, a standard pattern (HMM model) stored in advance in the comparison unit 13 is used. 1 to HMM model n) are compared (step S47). The result obtained in step S47 is output to an output stage (not shown) of the speech recognition apparatus 1 (step S48), and the series of speech recognition processing is terminated.

As described above, according to the speech recognition apparatus in the present embodiment, when extracting a speech feature amount by setting a time window from an analysis frame to be subjected to speech analysis, the power system feature parameter has the same period. The window length of the time window p (i) for extracting is set shorter than the time window for extracting the frequency axis system characteristic parameter.
Therefore, the dynamic feature (Δp (i)) of the power system feature parameter at the same time can be obtained with finer responsiveness while extracting the frequency axis feature parameter with a window length equal to or longer than the period of the basic pitch component in the vowel part. Therefore, the voice recognition rate is improved.

  In this embodiment, (f (i), Δp (i), Δf (i), ΔΔp (i)) is used as the feature vector sequence. However, consonant information of a human utterance is not included. A band-pass filter that passes only a band of 3 kHz to 8 kHz that is included in a large amount is provided, a short-time power-only feature vector Δtp (i) is extracted from audio data that has passed through the band-pass filter with a short window length pL, and a feature vector sequence (F (i), Δp (i), Δf (i), ΔΔp (i), Δtp (i)) may be used.

1 is a block diagram of a speech recognition apparatus according to an embodiment of the present invention. The figure which shows the mode of the setting of the time window with respect to the input audio | voice waveform in a time window position setting part. The figure which shows the characteristic parameter used for the speech recognition apparatus which concerns on embodiment of this invention. The flowchart which shows operation | movement of the speech recognition apparatus which concerns on embodiment of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Speech recognition apparatus 11 Time window position setting part 12 Voice feature-value extraction part 13 Comparison part 141-14n HMM model

Claims (6)

A time window of a predetermined length is set with a predetermined period for the analysis target voice, and this time window is used as a processing unit and includes a frequency axis characteristic parameter related to the frequency of the voice and a power system characteristic parameter related to the amplitude of the voice. A speech recognition apparatus comprising: a feature amount extraction unit that extracts a feature amount; and a speech recognition unit that recognizes the analysis target speech based on the feature amount extracted by the feature amount extraction unit,
The feature quantity extraction unit extracts the feature quantity by shortening a length of a time window for extracting only the power system feature parameter as compared with a length of a time window for extracting only the frequency axis system feature parameter. A voice recognition device characterized by the above.   The speech recognition apparatus according to claim 1, wherein the length of the time window for extracting only the frequency axis system characteristic parameter is equal to or longer than the period of the basic pitch component in the vowel part of the input speech.   The feature amount extraction unit calculates a difference feature parameter indicating a difference between temporally adjacent feature parameters among the feature parameters sequentially generated at the predetermined period, and supplies the difference feature parameter to the voice recognition unit. The speech recognition apparatus according to claim 1.   The feature amount extraction unit further includes a band pass filter unit that passes only a band of 3 kHz to 8 kHz with respect to the analysis target voice, and sets a power system characteristic parameter of the input voice via the band pass filter unit as the predetermined parameter. 2. The speech recognition apparatus according to claim 1, wherein the speech recognition apparatus sequentially extracts in a cycle and supplies the extracted difference between power system characteristic parameters adjacent in time to the speech recognition means. A time window of a predetermined length is set with a predetermined period for the analysis target voice, and this time window is used as a processing unit, and includes a frequency axis characteristic parameter related to the frequency of the voice and a power system characteristic parameter related to the amplitude of the voice. A speech recognition method for extracting a feature amount and recognizing the analysis target speech based on the extracted feature amount,
A speech recognition method for extracting the feature quantity by shortening a length of a time window for extracting only the power system feature parameter as compared with a length of a time window for extracting only the frequency axis system feature parameter . A time window of a predetermined length is set with a predetermined period for the analysis target voice, and this time window is used as a processing unit, and includes a frequency axis characteristic parameter related to the frequency of the voice and a power system characteristic parameter related to the amplitude of the voice. A speech recognition processing program comprising: a feature amount extracting step for extracting a feature amount; and a speech recognition step for recognizing the analysis target speech based on the extracted feature amount,
A speech recognition process for extracting the feature quantity by shortening a length of a time window for extracting only the power system feature parameter as compared with a length of a time window for extracting only the frequency axis system feature parameter program.

Images