JPS58150997A - Speech feature extractor - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a pronunciation feature extraction device that recognizes pronunciation from information other than speech.

Speech is converted into voice when the exhaled airflow sent out from the lungs passes through the vocal cords in the larynx, which vibrate, and is modulated by changing the shape of the exhaled air passage leading to the lips and nasal cavity. A voice is produced as a result of the comprehensive movement of the vocal tube.

Conventionally, in order to extract such sounds, the sound waves are converted into electrical signals using an acoustic microphone, which is then input to a number of filter circuits with predetermined frequency bands, and the output of each filter circuit is used to determine all the characteristics of the pronunciation. I was wearing it.

However, it is extremely difficult to perform speech recognition by extracting the pronunciation characteristics of all phonemes only from speech sounds and speech waves, which are the result of the comprehensive movement of the vocal tube.

, especially for non-stationary consonants, the noise energy 1[- is strong, and with the exception of voiceless fricatives /s, f/, etc., whose features can be almost certainly extracted in the speech wave, voiceless fricatives /h/
ya voiceless plosive/p+t, ni/ya voiced plosive/b, d
, g/ and nasal sounds /m, n, η/, etc. are extremely difficult to detect and separate.

In view of the above-mentioned drawbacks, the present invention has been developed by installing or arranging detectors that detect all movements of each part of the vocal tube in the vicinity of each part of the vocal tube, and having the outputs from each of the detectors processed by a processing device. An object of the present invention is to provide a pronunciation feature extraction device that can extract pronunciation more accurately than before.

Hereinafter, an embodiment of the invention will be described with reference to the drawings.

FIG. 1 shows a block configuration of a pronunciation extraction device according to an embodiment of the invention. In the figure, 1 is a vocal cord vibration detector attached near the vocal cords of the larynx to detect vibrations of the vocal cords; 2 is a wing vibration detector attached near the center of Makabe A4 to detect voice vibrations in the nasal cavity; 3 4 is a loss flow detector placed at the front of the oral cavity to detect loss flow; 4 is a palate contact detector placed on the palate in the oral cavity to detect contact between the tongue and the palate. 6 is a vocal cord vibration detector 1. This processing device extracts pronunciation characteristics from the outputs of the nasal vibration detector 22, the mouth airflow detector 3, and the palate contact detector 4. The configuration of the processing device 6 will be further explained in detail below with reference to FIG.

In FIG. 2, 6 is a threshold circuit that determines the presence or absence of vocal fold vibration based on a specific value from the vocal fold vibration information of the vocal fold vibration detector 1, and 6 is a threshold circuit that determines the presence or absence of vocal fold vibration based on a specific value from the nasal vibration information of the wing vibration detector 2. 8 is a differentiation circuit that determines the rate of change (acceleration) of the loss flow by differentiating the loss flow information from the loss flow detector 3; 9 is a differentiation circuit that determines the presence or absence of the rate of change of the loss flow; 1o is a threshold circuit that determines the presence or absence of a loss flow based on a specific value from the loss flow information of the airflow detector 3; 11 is a palate of the palate contact detector 4; 13 is a threshold circuit 6.7 which converts the contact information into a contact signal between the tongue and the roof of the mouth by the measuring circuit 12, and then judges all three types of states: pre-closing, retro-closing, and no closure, which will be described later. , 9.10, and three types of information in the tongue closure detection circuit 11.

A specific method of using the pronunciation feature extracting device configured as described above will be explained below with reference to FIG.

As shown in FIG. 3, the vocal cord vibration detecting section 1 detects vocal cord vibration by attaching an acceleration sensor 1' to the vocal cord part of the larynx in the human body using double-sided medical tape. The detected vocal fold vibration is output to the threshold circuit 6, and the threshold circuit 6 sends a positive (+) signal to the phoneme classification circuit 13 if the value of the vocal fold vibration is above a certain value; If so, output the signal.

Nasal vibrations are detected by attaching an acceleration sensor 2' as a folded blade vibration detector 2 to the center of the back wall of the human body using double-sided medical tape.

The detected nasal vibration is output to a threshold circuit 7, and the threshold circuit 7 detects the phoneme f-t if the value of the nasal vibration is above a specific value. By fixing and placing it on a desk, etc. in front of the oral cavity of the human body,
Detect loss flow. The detected loss flow is transferred to the differentiator circuit 8.
The differential circuit 8 calculates the rate of change of the loss flow and outputs the rate of change to the threshold circuit 9. Then, the threshold circuit 9 outputs a signal sound (g) to the phoneme classification circuit 13 if the value of the rate of change is above a specific value, and a signal (→signal) if the rate of change is below a certain value.On the other hand, the hot wire flowmeter The loss flow detected by the sensor 3' is also output to the threshold circuit 10, and in the threshold circuit 10, if the value of the loss flow is greater than or equal to a specific value, the loss flow is output to the phoneme classification circuit 1.
3 with a presence (-1-) signal, ! , if the value is below a certain value, no signal is output.

Further, as the palate contact detector 4, a contact sensor 4' as shown in FIG. 4 is used. The contact sensor 14' has a large number of electrodes 4'ak on the part that comes into contact with the tongue, and is attached to the upper part of the oral cavity in the human body by the stopper part 4'b, and the contact sensor 14' is connected to the tongue by the electrode 4'a. Detect contact status.

The detected contact state between the electrode 4/a and the tongue is sequentially input to the measurement circuit 12 and the tongue closure detection circuit 11, and when the contact state becomes a pattern as shown in FIG. When the information regarding closure becomes a pattern as shown in FIG. 6 (b), information regarding the latter closure is output to the phoneme classification circuit 13, and when there is no contact with the tongue, information indicating no closure is output to the phoneme classification circuit 13. be done.

Finally, the phoneme classification circuit 13 can determine the speech based on the information inputted from the threshold circuits 6, 7, 9° 1Q and the tongue closure detection circuit 11 from an internal memory table as shown in the table below.

Now, for example, if we have a phoneme wave as shown in FIG. 6 (a) [h? When the entire voice Ln2L-1 is uttered, the acceleration sensor 1' outputs a waveform as shown in FIG. 6(b) to the threshold circuit 6. The threshold circuit 6 then outputs a null (→ signal) for the "h" part and a present (10) signal for the "n" part to the phoneme classification circuit 13 based on the judgment based on the specific threshold value.

Furthermore, the acceleration sensor 1' outputs a waveform as shown in FIG. 6(c) to the threshold circuit 7. Then, the threshold circuit 7 outputs a null (-) signal at the section "rh" judged from a specific threshold value.
At the part "n", a presence (+) signal is output to the phoneme classification circuit 13.

Furthermore, the hot wire flow meter sensor 3' outputs a waveform as shown in FIG. 6 to the differentiating circuit 8 and the threshold circuit 10. Then, the threshold circuit 9 judges the differential value from the differentiator circuit 8 based on a specific threshold value and outputs a null (-) signal to the phoneme classification circuit 13 at the portions of "rh" and rnJ.

The threshold circuit 10 also outputs a presence (+) signal at the rh4 portion and a no (-) signal at the “n” portion to the phoneme classification circuit 13 based on a judgment based on a specific threshold value.

On the other hand, the contact sensor 4' detects the contact state between the electrode 4a and the tongue, and outputs the detected state to the tongue closure detection circuit 11 via the measurement circuit 12. Then, the tongue closure detection circuit 11 outputs information of "no closure" at the "h" portion according to the contact pattern, and outputs information of "anterior closure" to the phoneme classification circuit 13 at the "n" portion.

Then, the phoneme classification circuit 13 can fully recognize "h" and rnJ from an internal storage table as shown in the table based on each information.

As mentioned above, the vocal fold vibration detector 1. The nasal vibration detector 21, the mouth airflow detector 3, and the palate contact detector 4 detect all movements of each vocal tube, and the processing device 6 detects the information from a pre-stored table based on the information detected by each detector. By determining specific phonemes, it is possible to accurately recognize speech, which has been difficult in the past.

As described above, the present invention combines vocal cord vibration information detected by a vocal cord vibration detector, intranasal vibration information detected by a nasal vibration detector, loss flow information detected by a loss flow detector, and a palate contact detector. By providing a processing device that selects a specific phoneme based on the information stored in the device based on the contact information between the tongue and the roof of the mouth detected by the device, it is possible to extract pronunciation more accurately from the vocal tube than before. It is possible, and its practical effects are similar to that of a dog.

[Brief explanation of drawings]

0 Fig. 1 is a block diagram of a pronunciation feature extraction device according to an embodiment of the present invention, Fig. 2 is a block diagram of a processing device in the pronunciation feature extraction device, and Fig. 3 shows an example of use of the pronunciation feature extraction device. 4 is a plan view of the contact sensor, FIG. 5 is a diagram showing a contact pattern between the tongue and the palate, and FIG. 6 is a waveform diagram of each detector. 1... Vocal fold vibration detector, 2... Nasal vibration detector, 3... Loss flow detector, 4...
Palatal contact detector, 5...processing device. Patent applicant Makoto Ishizaka, Director of the Agency of Industrial Science and Technology - No. 3
Figure 4 Figure 5

Claims (1)

[Scope of Claims] A tactile detector attached to the larynx to detect vibrations of the vocal cords, the vocal cord vibration detector, and a nasal vibration detector. A pronunciation feature extraction device comprising a processing device that selects a specific phoneme from stored information based on the outputs of a loss flow detector and a palate contact detector.