JPH0243893A - voice recognition device - 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] Five hill division disaster The present invention relates to a speech recognition device.

BACKGROUND OF THE INVENTION In recent years, speech recognition devices have been actively developed, and speaker-specific and speaker-independent systems have been put into practical use. but,
The recognition rate of this device also changes significantly when the environment in which it is used changes.

For example, when sound is collected with a reflector such as a conference table or blackboard placed near a microphone, the transmission frequency characteristic will have a comb-shaped shape with many dips, and the speaker - microphone - reflection. Body position'+? ! Level characteristics vary greatly depending on the relationship. This is one of the factors that reduce the recognition rate when performing speech recognition by DP matching or the like. To examine the influence of transmission frequency characteristics on recognition performance, we conducted a recognition evaluation experiment using similar words by varying the distance between the speaker and the microphone for sound collection, and targeting speech with and without a reflector. The results are in the Proceedings of the Acoustical Society of Japan, March 1988, 269, 27.
Reported on page 0.

According to the report, three microphones placed at different distances from the speaker's mouth were used to measure the height from one reflector to the sound source (the speaker) and the height from the reflector to the microphone, respectively. When assuming 30Gm and the distance from the sound source to the microphone is 10°50.90■.

■The recognition rate when there is no reflector and there are no individual variations of the speaker (speech fluctuations that occur each time he/she speaks) is:
There is almost no difference due to the difference in microphone collection distance.

■When there is intra-individual variation, the recognition rate fluctuates by about 20%.

■The recognition rate when there is a reflector decreases as the distance between the speaker and the sound collection microphone increases, especially when L = 90.
This becomes noticeable in the case of an. Note that the value changes depending on the speaker.

(2) Even when there is a reflector, the recognition rate for L=50ao is within the range included in the amount of individual white variation of the speaker, but it is lower by several percent than when L=10a++ under the same conditions.

It is said that this was found out.

In this way, there is a difference in the recognition rate depending on whether there is a reflector nearby or not. This becomes noticeable, for example, when a speaker-independent recognition device is used in a car. With the speaker-specific method, it is possible to avoid this to some extent by creating a standard pattern for the usage environment, but with unspecified speakers, it is not possible to devise countermeasures against this problem because it is not known in what environment it will be used. . However, in a small space like the inside of a car, there are many things that reflect the sound, which changes the frequency characteristics of the microphone used, reducing the recognition rate.

Purpose The present invention was made in view of the above-mentioned circumstances.
In particular, the purpose of this invention is to provide a recognition device whose recognition rate does not decrease even in places with different acoustic characteristics, such as inside a car.

To achieve the above object, the present invention includes an acoustic/electrical converter that converts audio into an electrical signal, a filter group that analyzes the electrical signal, and a pattern comparison section that compares the analyzed results. In the speech recognition device, means for reproducing one or more tones of the center frequency of each filter, means for calculating and storing the sum or average value of the output of each filter, and the output value of each filter and the stored The present invention is characterized by comprising means for determining the difference between the two values, and means for increasing or decreasing the filter output depending on the magnitude of the difference. The following will explain one embodiment of the present invention.

FIG. 1 is a block diagram for explaining one embodiment of the present invention, in which 1 is a ROM, 2 is a D/A conversion circuit, 3 is an amplifier, 4 is a speaker, 5 is a microphone, and 6 is a microphone. amplifier, 7 is a filter group, 8 is a subtraction section, 9 is a register, 1
0 is an adder, 11 is a bit shift section, 12 is a register,
13 is a difference calculation section, 14 is a switch, 15 is a comparison section, 16
17 is a speech dictionary, 17 is a maximum similarity calculation unit, and 18 is a recognition result output unit. In the illustrated embodiment, the number of filters in the filter group 7 is 8, and the sound from the microphone 5 for voice input is input to the microphone amplifier 6. It is amplified and analyzed by filter group 7. The analyzed results are rectified and quantized by an A/D converter (not shown), and then put into the subtraction section 8 and stored in the register 9.0Although not shown in the diagram, the output of the microphone amplifier 6 is converted into a logarithm. is normal. As shown in FIG. 2, the subtraction unit 8
A predetermined value is subtracted from the output of each filter 71, 7□..., and the threshold value for subtraction is 8□8□'.
... are given to the subtraction units 81, 82, respectively, but
0 is entered as the initial value. Next, the eight values stored in the register 9 are added in the adder circuit 10, and when the bits are shifted three times in the bit shift section 11, the result becomes 1/8, and the average value is calculated, and this is sent to the register 12. First, turn switch 14 to side A, and register 9.
Subtract each value of 1 to 8 from the average value of register 12,
The value is substituted into the filter output subtraction unit 8, that is, the threshold value 81182/... in FIG. Therefore, the value Yi set to the threshold i is the eight values X1 (i=1 to
8) is expressed as Yi=X−Xi (1) using the average X. On the other hand, ROM1 has filter 7□,
A signal obtained by adding the sine waves of each center frequency of 7□ . . . 71 is encoded and stored in, for example, PCM. \This signal is converted to analog, amplified, and reproduced from the electroacoustic transducer (speaker) 4. At this time, amplitude values must be set for each frequency component stored in the ROMI so that the output level will be constant when reproduced. Perform the above adjustments while playing this sound.

FIG. 3 is a diagram showing the state of the above adjustment, in which the horizontal axis represents the channel number of each bandpass filter representing the frequency, and the vertical axis represents the level. Figure (a) shows the characteristics in free space, and speech dictionaries for speaker-independent recognition are created under these conditions. When this recognition device is brought into a narrow space such as a car (b)
The characteristics will be as shown in the figure. If we calculate the average level from these eight points, it will look like the broken line in the figure.

Furthermore, by subtracting each value in FIG. 2B from the average value according to equation (1), the result shown in FIG. 2C is obtained, and this value is used as each threshold value in FIG. After this adjustment, the output of each filter becomes as shown in Fig. (d), which can be corrected to the original characteristic as shown in Fig. (a), thereby preventing a decrease in recognition rate due to use in a narrow space. At the time of recognition, the switch 14 is turned to the B side, and the recognition is performed with the frequency characteristics corrected. Note that the figure shows the recognition part as the comparison part and the part that calculates the maximum similarity, but this is a necessary part regardless of the pattern matching method. Any method may be used, such as what is known as DP matching.

Effects As is clear from the above explanation, according to the present invention, even in a limited space such as a room, it is possible to correct the influence of sound reflection and flatten the frequency characteristics of the input sound from the microphone. As a result, the recognition rate can be improved.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram for explaining one embodiment of the present invention,
FIG. 2 is a detailed view of the subtraction section shown in FIG. 1, and FIG. 3 is the same for explaining the operation of the present invention. 1...ROM, 2...D/A conversion circuit, 3...amplifier, 4...speaker, 5...microphone, 6
...Microphone amplifier, 7.. Filter group, 8.. Subtraction section, 9.. Register, 10.. Adder, 11..
・Bit shift section. 12...Register, 13...Difference calculation unit, 14...
Switch, 15... Comparison section, 16... Voice dictionary, 1
7... Maximum similarity calculation unit, 18... Recognition result output unit. Figure 1

Claims (1)

[Claims] 1. In a speech recognition device that has an acoustic/electrical converter that converts speech into an electrical signal, a filter group that analyzes the electrical signal, and a pattern comparison section that compares the analyzed results, the center frequency of each filter is means for reproducing one or more sounds; means for calculating and storing the summation or average value of the outputs of each filter; and means for determining the difference between the output value of each filter and the stored value; 1. A speech recognition device comprising means for increasing or decreasing the filter output depending on the magnitude of the filter output.