JP2789158B2 - Voice recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition device, and more particularly to a speech recognition device capable of recognizing time-series data such as speech recognition.

[0002]

2. Description of the Related Art Conventional speech recognition apparatuses have attempted to recognize continuous speech by acoustic analysis. However, 100% of continuous speech cannot be recognized only by acoustic analysis. This can be seen from the fact that humans do not recognize speech only by auditory analysis. For this reason, research is being conducted to apply and use the fact that humans understand speech as a meaningful language to speech recognition devices.

[0003] In other words, the uncertainty of the acoustic analysis is analyzed by using the linguistic analysis, and the uncertainty of the lexical analysis is analyzed by using the linguistic analysis, taking into account the constraints of the speech expressing the words and sentences. Speech recognition has been attempted using However, a speech recognition device using this recognition method needs to develop artificial intelligence equivalent to the human brain.

[0004] Therefore, the present invention, with respect to the limit of acoustic analysis of speech, recognizes speech when a human utters a speech, based on the movement characteristics of articulators such as a mouth and a tongue that specify the speech. An object of the present invention is to provide a voice recognition device that can perform the voice recognition.

[0005]

[0006]

[0007]

A speech recognition apparatus according to the present invention comprises: a motion trajectory estimating means for estimating a first motion trajectory of an articulator based on an input first acoustic signal; A transit point estimating means for estimating a transit point of the first motion trajectory based on the first motion trajectory estimated by the means; And a voice generating means for generating a second acoustic signal based on a via point estimated by the via point estimating means or a phoneme string generated by the phoneme corresponding means. The generated second acoustic signal is compared with the first acoustic signal, and the transit point is estimated by the transit point estimating means until the difference between the second audio signal and the first acoustic signal becomes smaller than a predetermined first threshold value. of Repeated generation of the sound signal, and outputs the phoneme string obtained as a result as the recognition result of the first acoustic signal.

[0008] Preferably, the via point estimating means is a via point extracting means for extracting via points from or near the first motion trajectory estimated by the motion trajectory estimating means, and is extracted by the via point extracting means. First motion trajectory generation means for generating a second motion trajectory of the articulator based on the waypoints, wherein the second motion trajectory generated by the first motion trajectory generation means is calculated by motion trajectory estimation means. Compared with the estimated first motion trajectory, the extraction of the via point by the via point extraction means and the second motion by the first motion trajectory generation means until the difference becomes smaller than a predetermined second threshold value. The generation of the trajectory is repeated, and the obtained waypoint is output.

Preferably, the voice generating means converts the phoneme sequence generated by the phoneme correspondence means into a transit point of the third motion trajectory of the articulator, and the transit point is converted by the transit point conversion means. Second motion trajectory generating means for generating a third motion trajectory based on the waypoint, and generating a second acoustic signal based on the third motion trajectory generated by the second motion trajectory generating means Sound signal generating means.

[0010]

The speech recognition apparatus according to the present invention estimates a motion trajectory of an articulator which is presumed to perform a motion for generating an input acoustic signal, and further extracts a feature point specifying the motion trajectory. Thereby, continuous speech can be recognized as a phoneme string corresponding to the feature point.

[0011]

FIG. 1 is a block diagram showing a state from when an acoustic signal is input to a speech recognition apparatus according to an embodiment of the present invention until speech recognition is performed as a phoneme string corresponding to the acoustic signal. is there.

Referring to FIG. 1, a speech recognition apparatus according to one embodiment includes a phoneme sequence extraction unit 20 for extracting a sequence of phonemes from an audio signal, and a speech generation unit 3 for generating an audio signal from the sequence of phonemes.
0 is included. The phoneme string extraction unit 20 includes an audio signal input device 1 to which an audio signal is input, a motion trajectory estimation device 2 for estimating a motion trajectory, a via point estimation device 3 for estimating a via point,
And a phoneme correspondence device 4 for converting the waypoints into phonemes. The waypoint estimating device 3 is connected to the waypoint extracting device 11 and the first
And a motion trajectory generation device 12. Motion trajectory estimation device 2
Are described in, for example, Katsuhiko Shirai and Masaaki Yoshida, "Estimation of Articulatory Parameters from Speech Waves," IEICE Transactions, Vol. J61
-A, No. 5, pp. As disclosed in 409-416 (1978), articulation parameters of a predetermined articulation model are estimated from acoustic signals by a model matching method. Here, the articulatory model expresses the vocal tract shape based on the structure of articulatory organs. For example, Shirai and Yoshida, "Modeling of articulatory mechanisms and estimation of articulatory parameters by nonlinear multiple regression analysis," IEICE Trans. J. 59-A, 8, p668
(1976). Since the acoustic parameter is output according to the articulation parameter, the acoustic parameter is given as a function using the articulation parameter as a variable. The sound parameter is a parameter obtained by equivalently converting a sound signal. The basic method of estimating articulation parameters is to estimate an articulation parameter that gives a given acoustic signal (actually converted to an acoustic parameter), and minimizes an evaluation function J expressed by the following equation (1). So that the articulation parameters are varied. ^{_{J = {y s -y (x}} )} T {y s -y (x)} ... (1) However, _{y s} acoustic parameters (N dimensional vector) obtained from the audio signal, x is articulatory parameters, y (X) is a function for estimating an acoustic parameter (N-dimensional vector) from the articulation parameter x. T indicates transposition. Thus, acoustic parameter y _s articulatory parameters for realizing x, i.e. motion trajectory is estimated.

FIG. 3 is a diagram showing a pre-processed audio signal input to the speech recognition apparatus according to one embodiment of the present invention.
FIG. 4 is a diagram showing the trajectories and the waypoints of each articulatory organ estimated by the waypoint estimation device constituting the speech recognition device according to the embodiment of the present invention, and FIG. It is a figure which shows the phoneme string as a result of which the speech was recognized by the phoneme string extraction part of the speech recognition device.

Hereinafter, the operation of the apparatus will be described with reference to FIGS. An audio signal is input to the audio signal input device 1, and preprocessing such as noise processing of the input data is performed. Then, with respect to the pre-processed sound signal as shown in FIG. 3, the motion trajectory estimating apparatus 2 moves each articulatory organ that is presumed to move to realize the pre-processed sound signal. Estimate the trajectory. Further, the estimated motion trajectory is input to the waypoint estimation device 3.

[0015] Therefore, the waypoint extraction device 11 extracts the waypoints necessary for generating the input motion trajectory from or near the motion trajectory, and the first motion trajectory generation device 12
Give to. The first motion trajectory generating device 12 generates a trajectory passing through the extracted waypoint.

The transit point estimating device 3 compares the generated motion trajectory with the input motion trajectory, and until the accuracy is sufficient, that is, until the difference becomes smaller than a predetermined threshold value. The extraction of the via points by the extraction device 11 and the generation of the motion trajectory by the first motion trajectory generation device 12 are repeatedly performed. Via points and motion trajectories obtained by this repetition are as shown in FIG. 4, for example. Here, the curve by TBY is the trajectory of the vertical movement of the back of the tongue (the center of the tongue) with respect to time. In addition, TTY,
The curves by JY and LLY are curves of the vertical movement of the tongue tip, jaw and lower lip with respect to time, respectively.

The waypoints estimated by the waypoint estimation device 3 in this way need to be converted into phonemes for recognizing speech, and are input to the phoneme correspondence device 4. As a result, the phoneme correspondence device 4 converts the waypoints into phonemes, whereby the phoneme sequence 5 as shown in FIG. 5 is identified.

Therefore, it should be possible to recognize this phoneme sequence 5 as an audio signal input to the audio signal input device 1. Conversely, an audio signal input to the audio signal input device 1 is generated from the phoneme sequence 5. If not, the phoneme sequence 5
Is not enough to be a recognition result of the input acoustic signal.

For this purpose, a speech generator 30 of the speech recognition device is provided. The voice generation unit 30 includes a conversion device 6 for converting a phoneme into a via point, a second motion trajectory generation device 7 for generating a motion trajectory, and an audio signal generation device 8 for generating an audio signal. Second motion trajectory generator 7
Includes a motion command conversion device 9 and a motion prediction device 10.

The transit point conversion device 6 converts the phoneme sequence 5 into a target sequence of transit points of the posture of each articulator, that is, a sequence of points to which the motion trajectory such as the tongue, jaw, and lip passes. The target waypoint sequence converted by the waypoint conversion device 6 is input to the second motion trajectory generation device 7.

The motion command converter 9 calculates a torque (motion command) from the motion trajectory. The motion prediction device 10 calculates a motion trajectory from the torque. For example, the motion command converter 9
Is "Transactions of the Institute of Electronics, Information and Communication Engineers D-II Vol.
-D-II No. 5pp. 991-999 1992
The motion prediction apparatus 10 can employ the forward dynamics model (FDM) shown in the same journal. When such a motion command conversion device 9 and a motion prediction device 10 are combined, a nonlinear optimization problem can be solved.
Therefore, the motion command conversion device 9 and the motion prediction device 10
Accordingly, the second motion trajectory generating device 7 generates the motion trajectory of each articulator using the torque change, the muscle tension change, the motion command change, or the square integration minimization over the motion time of the motion command.

In order to generate the motion trajectory, as a method other than the above, the square integral minimization of the jerk (time derivative of acceleration) of the motion trajectory over the motion time by an approximate calculation using a spline function is used. You can also. This second
The same applies to the first motion trajectory generator 12 regarding the motion trajectory generator 7.

The motion trajectory of the articulator thus generated is input to an acoustic signal (voice waveform) generator 8. Then, the sound signal generation device 8 generates a sound signal from the input motion trajectory. Therefore, the sound signal generated by the sound signal generation device 8 and the sound signal input to the sound signal input device 1 can be compared. If the accuracy of this comparison is sufficient, the phoneme string 5 itself is used as the speech recognition result.

However, if the accuracy is not sufficient, the waypoint estimating device 3 estimates the waypoint again, and the sound signal generation via the waypoint conversion device 6 and the second motion trajectory generating device 7. Until the sound signal generation by the device 8 becomes sufficiently accurate, that is, the result of comparing the sound signal generated by the sound signal generation device 8 with the sound signal input to the sound signal input device 1, the difference is determined by a predetermined threshold value. Until it becomes smaller.

In this way, the speech recognition device can recognize the phoneme sequence when the accuracy becomes sufficient as the phoneme sequence corresponding to the audio signal input to the audio signal input device 1.

FIG. 2 is a block diagram showing a state from when an acoustic signal is input to a speech recognition apparatus according to another embodiment of the present invention to when speech is recognized as a phoneme sequence corresponding to the acoustic signal. It is.

In the embodiment shown in FIG. 1, the phoneme string extracting unit 2
The transit point estimated by the transit point estimating device 3 constituting the zero is converted into a phoneme once by the phoneme correspondence device 4, and the phoneme sequence 5 which is a sequence of the phonemes is transmitted to the transit point constituting the voice generating unit 30. The conversion device 6 converts the articulatory organ's posture into a target via-point sequence, and using this target via-point sequence, the second motion trajectory generating device 7 generates a motion trajectory.

On the other hand, in this embodiment, the via point estimation device 3 and the second motion trajectory generation device 7 are connected as shown in FIG. In this case, the waypoints are not converted to phonemes,
As in the embodiment shown in FIG. 1, it is possible to compare the acoustic signal generated by the acoustic signal generating device 8 with the signal input to the acoustic signal input device 1.

However, it is an object of the present invention to recognize an input audio signal as a phoneme sequence even if the audio signals can be compared.

For this reason, the point at which the sound signal generated by the sound signal generation device 8 and the sound signal input to the sound signal input device 1 have sufficient accuracy is converted into a phoneme by the phoneme correspondence device 4. I do.

Thus, as in the embodiment shown in FIG. 1, the acoustic signal input to the acoustic signal input device 1 can be recognized as a phoneme sequence.

[0032]

As described above, according to the present invention, the sound signal generated by the movement of the articulator is converted into a characteristic point for specifying the movement of the articulator, and is extracted. It can be recognized as a corresponding phoneme sequence.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a state from when an acoustic signal is input to a speech recognition apparatus according to an embodiment of the present invention until speech recognition is performed as a phoneme sequence corresponding to the acoustic signal.

FIG. 2 is a block diagram showing a state from when an acoustic signal is input to a speech recognition apparatus according to another embodiment of the present invention to when speech is recognized as a phoneme sequence corresponding to the acoustic signal.

FIG. 3 is a diagram showing a pre-processed audio signal input to the speech recognition apparatus according to the embodiment of the present invention;

FIG. 4 is a diagram showing a trajectory and a waypoint of each articulatory organ estimated by a waypoint estimation device constituting the speech recognition device according to one embodiment of the present invention;

FIG. 5 is a diagram showing a phoneme sequence as a result of speech recognition by the phoneme sequence extraction unit of the speech recognition device according to one embodiment of the present invention.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 acoustic signal (voice waveform) input device 2 motion trajectory estimation device 3 via point estimation device 4 phoneme correspondence device 5 phoneme sequence 6 conversion device to via point 7 second motion trajectory generation device 8 acoustic signal (voice waveform) generation device 9 Motion command conversion device 10 Motion prediction device 11 Via point extraction device 12 First motion trajectory generation device 20 Phoneme string extraction unit 30 Voice generation unit

──────────────────────────────────────────────────続 き Continuation of the front page (56) References Japanese Patent Publication No. 52-14562 (JP, B2) Transactions of the Institute of Electronics, Information and Communication Engineers (May 1978) Vol. J61-A, no. 5, p. 409-416 (58) Fields surveyed (Int. Cl. ⁶ , DB name) G10L 9/10 301 JICST file (JOIS)

Claims (3)

(57) [Claims] 1. A motion trajectory estimating means for estimating a first motion trajectory of an articulator based on an input first acoustic signal; and a first motion trajectory estimated by the motion trajectory estimating means. Via point estimating means for estimating a via point of the first motion trajectory; phoneme correspondence means for converting each via point estimated by the via point estimating means into a phoneme to generate a phoneme sequence; Voice generating means for generating a second audio signal based on the waypoint estimated by the means or the phoneme sequence generated by the phoneme correspondence means, and the second audio signal generated by the voice generating means Compared with the first sound signal, the estimation of the waypoint by the waypoint estimation means and the generation of the second sound signal by the sound generation means until the difference becomes smaller than a predetermined first threshold value. Repeat and, the resulting phoneme sequence the first
A speech recognition device that outputs a sound signal as a recognition result. 2. The transit point estimating means includes: transit point extracting means for extracting the transit point from or near the first motion trajectory estimated by the motion trajectory estimating means; And a first motion trajectory generating means for generating a second motion trajectory of the articulator based on the obtained waypoint, wherein the second motion trajectory generated by the first motion trajectory generating means performs the motion Comparing with the first motion trajectory estimated by the trajectory estimating means, extracting the waypoint by the waypoint extracting means and generating the first motion trajectory until the difference becomes smaller than a predetermined second threshold value 2. The speech recognition device according to claim 1, wherein the generation of the second motion trajectory by the means is repeated, and the obtained waypoint is output. 3. The transit point converting means for converting a phoneme sequence generated by the phoneme correspondence means into a transit point of a third motion trajectory of the articulator, wherein Second motion trajectory generating means for generating the third motion trajectory based on the obtained waypoint, and the second sound based on the third motion trajectory generated by the second motion trajectory generating means The speech recognition device according to claim 1, further comprising: an acoustic signal generation unit configured to generate a signal.