JPH0511798A - Voice recognition device - Google Patents

Abstract

PURPOSE:To eliminate the need for all high order parameters at real times and for special hardware which performs high order parameter analysis at a high speed. CONSTITUTION:A voice recognition device consists of a low order parameter analysis section 2 which finds the low order parameter time series of input voice, a leading and trailing end detection section 3 which detects the leading and trailing ends of input voice by using the parameter time series found by the low order parameter analysis section 2, a high order parameter analysis section 5 which finds high order parameters at preset sample points from within input voice corresponding to the detected leading and trailing ends, a pattern collation section 7 which collates featured parameters by the found high order parameters with previously registered reference parameters to calculate similarity and a recognition result output section 8 which outputs the recognition result of input voice in accordance with the calculated similarity.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device for recognizing a voice by extracting a parameter from an input voice.

[0002]

2. Description of the Related Art In general speech recognition methods, a method in which a low-order parameter and a high-order parameter are respectively extracted from an input voice and the input voice is accurately recognized using both of them is the mainstream. This is a method of simultaneously performing low-order parameter analysis and high-order parameter analysis on the input voice in real time. However, in the high-order parameter analysis, since it is necessary to analyze the DFT spectrum, filter bank, LPC parameters, etc. in real time every 8 msec, special high-speed signal processing hardware such as DSP (digital signal processor) is required.

For example, according to a conventional acoustic recognition system described in "The Recognition of Telephone Speech by Hybrid Structure Matching Method" (Asada et al.) In March 1982, the Acoustic Society, first, the input speech (the output of the speech input section) Low order parameter analysis and high order parameter analysis are performed in parallel. The start and end of the input voice are detected according to the result of the low-order parameter analysis, and the resample frame is determined from the range of the input voice corresponding to the start and end according to a preset number of frames. According to this resample frame number, a fixed-dimension speech feature vector for a high-order parameter is extracted from the high-order parameter analysis results (which have already been obtained) corresponding to the range of the input speech, and the standard pattern of the registered word dictionary is extracted. To calculate the degree of similarity. The recognition result is output according to the similarity.

[0004]

In the conventional speech recognition method as described above, first, all the input voices are subjected to high-order parameter analysis, and the word feature vector is extracted from only the necessary high-order parameters. . That is, since only a few frames of the total high-order parameter analysis result of the input voice are used, most of the high-order parameters calculated at all are wasted. In other words, the calculation unnecessary for recognition is performed.

As described above, in the conventional speech recognition method, special hardware is required to analyze the high-order parameters in real time for all input voices, and most of the calculated high-order parameters are wasted. Become.

An object of the present invention is to provide a speech recognition apparatus which does not require special hardware by analyzing the high order parameters of the input speech only for the necessary frames.

[0007]

According to the present invention, a low-order parameter analysis section for obtaining a low-order parameter time series of input speech, and a parameter order series obtained by the low-order parameter analysis section A start-end detection unit that detects the start-end, a high-order parameter analysis unit that obtains a high-order parameter at a predetermined sample point from the range of the input voice that corresponds to the start-end detected by the start-end detection unit, and this high-order parameter analysis unit. A pattern matching unit that calculates the similarity by matching the feature parameter based on the parameters obtained by the order parameter analysis unit with the standard parameter registered in advance, and the recognition result for the input voice according to the similarity calculated by the pattern matching unit. There is provided a voice recognition device including a recognition result output unit for outputting

[0008]

In the present invention, since all high-order parameters are not calculated in real time, special hardware such as DSP for performing high-order parameter analysis at high speed is not required. Therefore, the voice recognition can be realized by the calculation function of a personal computer.

Further, even when the calculation function is large like a high-speed workstation and real-time processing of high-order parameters is possible only by software, multitask processing is performed and other tasks simultaneously run in parallel. In such a system, the calculation efficiency of other tasks is increased because unnecessary high-order parameters are not calculated.

[0010]

Embodiments will be described below with reference to the drawings.

In the voice recognition device according to the embodiment of the present invention shown in FIG. 1, the voice input unit 1 converts a voice signal input via a microphone or the like into a digital signal. The voice input unit 1 has, for example, a low-pass filter with a cutoff frequency of 5.6 kHz for removing high-frequency noise components included in the input voice signal, and an input voice (analog signal) captured through the low-pass filter, for example, with a sampling frequency of 12 kHz. , And an A / D converter for converting into a digital signal with a quantization bit number of 12 bits.

The low-order parameter analysis unit 2 uses energy,
For example, an analysis window length of 24 mse is used as a low order parameter for obtaining voice data by reducing the number of zero crossings.
c, Calculation is performed every 8 msec of the analysis frame period. The obtained low-order parameter is used by the next start / end detection unit 3 to detect the start / end of a word.

As shown in FIG. 2, the start / end detection unit 3 sets a threshold value for energy, which is a low-order parameter, and defines a time point at which the energy of the analysis frame becomes larger than this threshold value as the start point of the word and the energy is smaller than the threshold value. The time when it becomes is the end of the word. Even if the number of zero crossings is used as a low-order parameter instead of energy, the start and end of a word can be detected in a similar manner. Various methods have been proposed in the past for detecting the start and end of voice, for example, Y. Takabayashi, H. Shinaoda,
H.Asada, T.Nitta, S.Hirai and S.watanabe, Teleph
one speech recognition using a hybrid method ", Pro
The start / end detection method disclosed in c. Seventh IICPR, pp.1232-1235, 1984 can be used.

The resample frame determination unit 4 determines a resample frame, that is, a frame for high order parameter analysis according to the detected start and end points. For example, the determination unit 4 determines, for example, 10 frames at equal intervals from the frames to be analyzed between the detected start and end points as shown in FIG. However, the number of frames is predetermined. Of course, the resample frames may be determined not at regular intervals but at irregular intervals in consideration of data and the like. Where 10
When frames are needed, the actual frame number of each of these frames is determined.

The high-order parameter analysis unit 5 requires many operations such as filter analysis output, DFT spectrum, cepstrum, and LPC parameters from the voice signal from the voice input unit 1 corresponding to only the determined resample frame. Find the parameters. Speech feature vector extraction unit 6
Calculates a word feature parameter from the obtained high-order parameter. For example, when using a 12th-order high-order parameter, a 12 × 10 = 120-dimensional word feature vector is obtained as follows.

When a voice feature vector is obtained by FFT (Fast Fourier Transform), for example, 24 mse
A frequency spectrum (DFT spectrum) Xk having a resolution of 128 points is obtained by performing FFT of 128 points for c time length. Power of this frequency spectrum Xk | Xk | ² Is smoothed in the frequency direction and is divided into 12 in the frequency direction, whereby 12 channel (dimensional) filter bank equivalent outputs Zi (i = 1, 2, ...
・, 12) is required. Specifically, a 12-channel filter bank equivalent output Zi (i = 1, 2, ... 1)
When seeking 2),

[0017]

[Equation 1] As a result, the frequency spectrum is smoothed in the frequency direction. This is obtained. Filter bank equivalent output Zi
By logarithmizing (i = 1, 2, ..., 12), a 12-dimensional high-order parameter is obtained as Gi = 10log Zi (i = 1, 2, ..., 12). Further, when the word feature vector is obtained by digital filter bank, LPC analysis or cepstrum analysis, the same method as above is executed.

The pattern matching unit 7 calculates the degree of similarity between the obtained word feature vector and the word standard feature vector of the registered voice word dictionary. For this similarity calculation, for example, collation using statistical pattern recognition such as Mahalanobis distance or Neulau network is possible in addition to the composite similarity method based on a fixed dimension feature vector. The recognition result is obtained by outputting from the recognition result output unit 8 the word having the highest similarity to the input voice among the similarities thus obtained.

According to the above-mentioned embodiment, since the useless calculation of the high-order parameter is eliminated, the total calculation amount is reduced. However,
As shown in FIG. 3, in the method of the present invention, since the high-order parameter analysis is performed after the start / end detection, a slight time delay occurs before the similarity calculation, which is the next process. However, since the number of sample points required for pattern matching can be reduced in the time direction (for example, up to 10 points), the time required for calculating the high-order parameter does not increase so much, which is particularly effective.

Normally, when determining the beginning of a voice, if the input voice energy level exceeds the threshold value E0 at time t0 and a certain period of time elapses, then the time t0 is determined as the beginning. When determining the end of the voice, the voice energy level becomes equal to or lower than the threshold value E0, and the preset period TQ
During the period, if the threshold value E0 or less is maintained, the time when the voice energy level reaches the threshold value E0 is determined as the end as shown in FIG. The method of determining the termination is described in, for example, J.
G. Wilpon, LR Rabiner, & T. Mortion: "AnImproved
Word-Detection Algorithm for Telephone-Quality Spe
ech Incorporating Both Syntactic and Semantic Con
straints "AT & T Bell Technical T.63.3, 479 (1984)
The method disclosed in can be used.

If the end of the voice is obtained and the high-order parameter analysis is performed as in the usual method, the high-order parameter analysis is always delayed by a fixed time (TQ).
Therefore, an embodiment in which the delay time is not included in the time required to obtain the recognition result is shown in FIG.

In this embodiment, a low-order parameter analysis unit 2 for analyzing a low-order parameter of the voice output of the voice input unit 1
Is connected to the first input terminal of the start / end candidate detection unit 3a. The start / end candidate detection unit 3a receives the output signal output from the low-order parameter analysis unit 2a and detects candidates that are the start and end of a word. Start / End candidate detector 3a
Is connected to the first input terminal of the high-order parameter analysis unit 5a via the resample candidate determination unit 4a. The second input terminal of the high-order parameter 5a is connected to the output terminal of the voice input unit 1. The output terminal of the high-order parameter analysis unit 5a and the second output terminal of the low-order parameter analysis unit 2a are connected to the first and second input terminals of the start / end determination unit 3b. The first output terminal of the start / end determination unit 3b is connected to the input terminal of the word feature vector extraction unit 6. Start and end determination unit 3
The second output terminal of b is connected to the second input terminal of the start / end candidate detection unit 3a. The word feature vector extraction unit 6 is connected to the recognition result output unit 8 via the pattern matching unit 7. Next, the operation of the voice recognition device shown in FIG. 5 will be described with reference to the flowchart of FIG.

When a voice signal corresponding to the voice "six", for example, is input from the voice input unit 1 to the low order parameter analysis unit 2a, the low order parameter analysis unit 2a inputs the input voice signal as in the embodiment of FIG. To analyze low order parameters. The analysis voice signal is input to the start / end candidate detection unit 3a, so that the start / end candidate detection unit 3a first outputs the voice (si
x) The starting point candidate of x) is detected. In this start point candidate detection,
When the input voice level exceeds the threshold value Eo at the time point t0 and a predetermined time TQ, for example, 0.2 seconds has elapsed, the time point t0 is detected as a start point candidate as shown in FIG.

Next, a termination candidate is detected. In this termination candidate detection, when the voice energy becomes equal to or lower than the threshold Eo and does not exceed the threshold Eo for a certain period of time TQ, the time when the voice energy reaches the threshold Eo as shown in FIG. 7 is detected as the termination candidate. It

When the start and end candidates detected by the start and end candidate detecting section 3a are input to the resample candidate determining section 4a, the resample candidate determining section 4a determines sample points in the range determined by the start and end candidates. To do. When the output signal of the resample candidate determination unit 4a is input to the high order parameter analysis unit 5a, the high order parameter analysis unit 5a.
Computes the high order parameters at the determined sample points.

The start / end determination unit 3b receives the output signals from the low-order parameter analysis unit 2a and the high-order parameter analysis unit 5a to determine whether or not the termination candidate is the termination. This can be realized by monitoring whether or not the voice energy after the termination candidate point is equal to or lower than the threshold value Eo and does not exceed the threshold value TQ for a certain period of time as shown in FIG. As shown in Fig. 8, "six" is uttered,
If the voice energy does not exceed the threshold value Eo for the fixed time length TQ or more, the start / end determination unit 3b determines that the input end candidate is not the end. At this time, the start / end determination unit 3b outputs the output signal to the start / end candidate detection unit 3a and the resample candidate determination unit 4
a and the high-order parameter analysis unit 5a. The high-order parameter analysis unit 5 interrupts the calculation of the high-order parameters and the like, and the start / end candidate detection unit 3a again obtains the start / end candidates. After this, the resample candidates are determined again and the high-order parameter analysis is performed.

When the start / end determination unit 3b determines that the end candidate is the end, the start / end determination unit 3b sends the speech energy in the range of the start / end to the word feature vector extraction unit 6. The word feature vector extraction unit 6 extracts a word feature vector from the sent voice energy. The extracted word feature vector is input to the pattern matching unit 7 and matched with the reference word feature vector. When the similarity obtained by this collation is sent to the recognition result output unit 8, the recognition result output unit 8 outputs the recognition result.

According to the embodiment shown in FIG. 1, there is always a time delay of a fixed time length TQ for detecting the start and end, but according to the embodiment shown in FIG. 5, the analysis process is started immediately after the end candidate is detected. The fixed time delay required for detecting the end of is not included in the time until the recognition result is obtained. Therefore, faster recognition processing can be realized while maintaining the recognition performance.

[0029]

As described above, according to the present invention, since all high-order parameters are not obtained in real time, special hardware such as DSP for performing high-order parameter analysis at high speed is not required. Therefore, voice recognition can be realized even with the calculation function of a small computer such as a personal computer.

Further, when a computer system such as a high-speed workstation, which has a large calculation function, an operating system such as VIX is installed, and which is capable of real-time processing of high-order parameters only by software, multitask processing is used. The low-order parameter analysis and the high-order parameter analysis can be simultaneously performed in parallel. In this case, according to the present invention, since unnecessary high-order parameters are not calculated, other tasks can be executed correspondingly, and the calculation efficiency for speech recognition is improved. By doing so, it is possible to efficiently use the resources of the computer that wasted by performing wasteful processing by voice recognition for other tasks.

[Brief description of drawings]

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

FIG. 2 is a diagram for explaining an example of start / end detection and voice feature vector extraction according to the present invention.

FIG. 3 is a diagram showing a processing time after the start and the end are detected by the voice recognition device of the present invention.

FIG. 4 is a diagram illustrating start / end detection.

FIG. 5 is a block diagram of a voice recognition device according to another embodiment.

FIG. 6 is a flowchart for explaining the operation of the voice recognition device in FIG.

FIG. 7 is a diagram for explaining an example of detecting start and end candidates.

FIG. 8 is a diagram illustrating an example of detection of start and end points and extraction of a voice feature spectrum.

[Explanation of symbols]

1 ... Voice input unit, 2 ... Low-order parameter analysis unit, 3 ...
Start-end detection unit, 4 ... Resample frame determination unit, 5 ... High-order parameter analysis unit, 6 ... Word feature vector extraction unit,
7 ... Pattern matching unit, 8 ... Recognition result output unit.

Claims (4)

[Claims] 1. Low-order parameter analysis means for obtaining a low-order parameter time series from input speech, and start / end detection means for detecting the start / end of the input speech using the parameter time series obtained by the low-order parameter. A high-order parameter analysis means for obtaining a high-order parameter at a predetermined sample point from the range of the input voice corresponding to the start and end detected by the start-and-end detection means, and the high-order parameter analysis means. A voice recognition device comprising pattern matching means for matching the parameter-based feature pattern with a standard pattern registered in advance to identify the input voice. 2. The voice recognition device according to claim 1, wherein the low-order parameter includes an analysis result such as voice power and the number of zero crossings with a small amount of calculation. 3. The speech recognition apparatus according to claim 1, wherein the high-order parameters include analysis results with a large amount of calculation such as DFT spectrum, filter bank analysis, and LPC parameters. 4. A low-order parameter time series is obtained from an input voice, and a low-order parameter analyzing means for outputting voice energy, and the input from the relation between the level of the voice energy output from the low-order parameter analyzing means and a threshold value. A start / end candidate detecting means for detecting a start / end candidate of voice, and a high order for obtaining a high order parameter at a predetermined sample point from the range of the input voice corresponding to the start / end candidate detected by the start / end detecting means. Parameter analysis means, a start / end determination means for monitoring the relationship between the level of the voice energy and a threshold value, and determining the start / end candidate as the start / end, and the high-order parameter analysis in response to the determination of the start / end. A voice recognition device comprising: identification means for identifying the input voice based on the parameter obtained by the means.