JPS59123898A - Tester for voice recognition equipment - 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 (5) Technical Field of the Invention The present invention provides a test circuit for a speech recognition device, particularly for the analog signal processing portion of the speech input section of the speech recognition device, which can be easily tested in one examination. The present invention relates to a speech recognition device test circuit that enables the following.

0 Technical background and issues In general, with any device9 testing at the time of design to see if it operates as expected, and testing and diagnosis after operation are two important things, but this also applies to speech recognition devices. , is similar.

A speech recognition device converts a voice input analog signal into a digital signal, analyzes it to extract feature parameters, and compares and matches the feature parameters with standard feature parameters registered in a dictionary in advance. It is designed to recognize the input voice. Traditionally, testing of speech recognition devices has focused on the digital signal processing part, such as checking whether the speech recognition results are correct, and little consideration has been given to tests focusing on the analog signal processing part. Ta. In addition, the device did not have a self-test function, and when conducting a test, nine people had to create and input test data one by one, and two people had to take the output from the test section and check it.

For example, when automating tasks such as checking bank deposit balances, it is being considered to use a multi-line voice recognition device connected to a telephone line. In this case, it is difficult to confirm the normality of the voice recognition device by actually inputting the voice from each line in the field, which can easily lead to operational errors during the test, and the test takes a long time. There are problems such as the need for experienced line engineers.

(G. Purpose and Structure of the Invention The present invention aims to solve the above-mentioned problems, and provides a self-test/diagnosis function using a simple test circuit. In particular, the test centering on the analog circuit of the audio input section can be easily performed. The purpose is to quickly perform the test.For this reason, the speech recognition device test and test circuit of the present invention are as follows.

A voice recognition device comprising at least an audio input section that inputs an audio input analog signal and performs band limiting, gain adjustment, and analog-to-digital conversion, and a power level calculation circuit that calculates an input power level from the output digital signal of the audio input section. In the device.

a memory circuit that stores test data regarding waveforms of multiple strokes having different frequencies and amplitudes; a digital-to-analog conversion circuit that reads out the data stored in the memory circuit and converts it into an analog signal;
a switching circuit that applies an output analog signal of the analog conversion circuit to the audio input section;

A checking section is provided for comparing the output data of the power level calculation circuit with a predetermined expected value corresponding to the data read from the storage circuit, and the frequency characteristics and gain characteristics of the audio input section are tested. This is expressed as a percentage. Embodiments will be described below with reference to the drawings.

Embodiment of the Invention FIG. 1 shows the configuration of an embodiment of the invention, FIG. 2 shows a waveform diagram of a test signal, and FIG. 3 shows an example of application of the invention to a multi-line speech recognition device.

In the figure, 1 is an audio input terminal, 2 is a switching circuit, 3 is an audio input section, 4 is a bandpass filter, 5 is a gain adjustment circuit, 6 is an analog/digital conversion circuit, 7 is an audio analysis section, and 8 is a frame division 9 is a feature parameter extraction unit, 10 is a power level calculation unit, 11 is a control unit, 12 is a speech recognition unit, 13i
, a main dictionary, 14 a storage circuit, and 15 a digital-to-analog conversion circuit.

16 represents a power level calculating section.

An analog signal of a voice to be recognized is input to the voice input terminal 1 . The switching circuit 2 is a circuit that switches between an input signal in the normal recognition mode and a test key signal in the test mode under the control of the control unit 11. In normal recognition mode.

The switch of the switching circuit 2 is set to the A side shown in the figure.

Audio input analog signal from audio input terminal 1.

The signal is supplied to the audio input section 3. In test mode,
The switch of the switching circuit 2 is set to the B side shown in FIG. 9, and a test signal from the digital-to-analog conversion circuit 15 is applied to the audio input section 3.

The audio input section 3 is mainly composed of circuits that process analog signals, and includes, for example, a bandpass filter 4. Gain adjustment circuit 5.
It has an analog-to-digital conversion circuit 6. The bandpass filter 4 has a frequency range from 300Hz to 3.2, for example, when the voice signal to be recognized is a signal input through a telephone line.
Removes unnecessary signals outside the telephone band at KHz t and performs band limiting. The gain adjustment circuit 5 is a circuit that amplifies the input signal to an appropriate strength. Masanaka Masanaka's analog signal is converted to 1, e.g.
．． It is sampled at 8 KHz and converted to a digital signal. This digital signal is sent to the voice analysis section 7.

Does the audio analysis unit 7 perform frame division? 'r1+s, feature parameter extraction unit 9. Power level - Mill: It has a mountain part 1o, etc. The frame division unit 8 divides the input audio converted into a digital signal into nine frames, each having an interval of, for example, 15m5. The power level calculation section 10 is.

The input power level of each frame divided by the frame dividing unit 8 is calculated. Also, the feature parameter extraction unit 9b2, for example, has a self-phase opening in the frame 4! Add a total of 3 l numbers,
Linear predictive analysis is performed to extract feature parameters representing the spectral envelope. These analysis results are sent to the control unit 1
1 will be notified. The control unit 1 controls the entire voice recognition equipment, and when it receives a notification from the voice analysis unit 7, for example, it detects a voice section from the input power level.

The speech recognition unit 12 is instructed to perform pattern matching.

In the dictionary 13, standard feature parameters are registered and stored in advance in correspondence with each recognized word. Speech recognition unit 1
2. The feature parameters extracted by the speech analysis unit 7,
The input speech and the registered words in the dictionary 13 are matched by sequentially calculating the distance from each standard feature parameter registered in the dictionary 11''13.

The control unit 11 determines the recognized word based on the matching result of the voice recognition unit 12, and recognizes the voice.

Memory circuit 14. The digital-to-analog conversion circuit 15 and the power level check section 16 are circuits provided for testing the speech recognition device.

The memory circuit 14 stores, for example, pulse code modulation (PCM) of a plurality of sine waves with different frequencies and amplitudes in advance.
Data is stored and stored. The digital-to-analog conversion circuit 15 reads out the PCM from the storage circuit 14.
This is a circuit that converts data into analog signals. In the test mode, the power level check section 16 checks the frequency characteristics and gain characteristics of the audio input section 30 by comparing the power level obtained in the power level calculation section 1o with a predetermined expected value. It is something to do.

For example, when the control unit 11 is notified of the test mode by a signal line or a switch (not shown), 9
It works like this: First, the control unit 11 connects the switching circuit 2 to the side B in the figure using the 2 selection signal, and also connects the switching circuit 2 to the side B in the figure using the readout signal.

The test data read out from the storage circuit 14 is supplied to the primary digital-to-analog conversion circuit 15. , converts the input test data into an analog signal, and sends it to the audio input unit 3 via the switching circuit 2. In other words, the audio input unit 3 has a frequency and amplitude signal as shown in FIG. 2, for example. The resulting intersecting number of sine waves are applied in a time-division manner.

Test data, for example as shown in FIG. Gain adjustment circuit 5. After passing through the analog/digital conversion circuit 6.

It is input to the voice analysis section 7. The frame division section 8 of the audio analysis section 7 divides the input signal into frames at predetermined time intervals, and the power level calculation section 10 calculates the input power level for each frame and notifies the control section 11 of the input power level. Control unit 11
For example, since it is the test mode, the input power level notified in step 2 is transmitted to the power level check section 16, and a check is instructed.

The power level check unit 16 compares the transmitted power level with a predetermined waiting value corresponding to the test data read out from the storage circuit 14.
Diagnose whether the frequency characteristics and gain characteristics are correct,
If a predetermined expected value is not obtained, an error message is displayed.

The present invention can be applied to a multi-use voice recognition device as shown in FIG. 3, for example.

In FIG. 3, numerals 2, 11.147 and 16 correspond to those in FIG. 1, 1-1 to 1-n are audio input terminals, 3-1 to 3-n are audio input units, and 7-1 to 7-n represents a speech analysis section.

The audio input terminals 1-1 to i-n are each connected to a different telephone line, and input audio from the telephone line can be input independently.
Voice input units 3-1 to 3-71 and voice analysis units 7-1 to
It is designed to be recognized via 7-71.

According to the speech recognition device shown in FIG. 3, simultaneous speech recognition of up to n lines is possible, but the memory circuit 14 for testing.
One set of the digital/analog circuit 15 and the power level check section 16 is sufficient. When conducting a test, the control unit 11
teeth.

Depending on the selection signal, the audio input units 3-1 to 3 to be tested
-17, and controls the switching circuit 2 to connect it to the output line of the digital/analog circuit 15. After the connection, the frequency characteristics and gain characteristics of the selected audio input section are checked in the same way as in case 197 shown in FIG.

Next, all audio input sections 3-1 to 3-
If this is repeated for n, one set of test circuits will perform the same tests as when the test was performed by sending test signals from all lines. Therefore, the amount of hardware per pi) may be small.

(According to the present invention, as described in detail, it is possible to efficiently test and diagnose a speech recognition device using a simple circuit, and it becomes easy to estimate and detect a failure location.) It is also possible to detect deterioration in the performance of the device at an early stage and prevent failures from occurring.For example, by installing a circuit that generates a test sound,
There is no need to conduct tests such as voice recognition, and the frequency characteristics and gain characteristics are almost automatically checked by inputting multiple simple sine waves in a time-sharing manner, making test operations easy and requiring only experienced engineers. There is no need for this, and there is almost no chance of operational errors. In addition, it is possible to quickly and reliably test the voice recognition system for 11 multiple times, and the reliability can be improved.

[Brief explanation of the drawing]

FIG. 1 shows an embodiment of the present invention, FIG. 2 shows a waveform diagram of a test signal, and FIG. 3 shows an example of application of the present invention to a multi-line speech recognition device. In the figure, 1 is an audio input terminal, 2 is a switching circuit, 3 is an audio input section, 7 is an audio analysis section, 10 is a power level calculation section, 11
14 represents a control unit, 14 represents a memory circuit, 15 represents a main digital-to-analog conversion circuit, and 16 represents a power level check unit. Patent applicant: Fujitsu Ltd. Representative Patent Attorney Mori 1) Hiroshi (and 1 other person) 5 ″i′z膓

Claims (1)

[Claims] An audio input section that inputs an audio input analog signal and performs band limiting, gain adjustment, and analog-to-digital conversion, and a power level calculation circuit that calculates an input power level from the output digital signal of the audio input section. At least a memory circuit for storing test data regarding the waveform of the initial seed a having two different frequencies and amplitudes in a speech recognition device equipped with the device, and a digital circuit for reading out the data stored in the memory circuit at around j1 and converting it into an analog signal. - An analog converter circuit, a switching circuit that applies the output analog signal of the digital-to-analog converter circuit to the audio input section, and a predetermined control circuit that reads out the output data of the power level calculation circuit from the memory circuit, and also outputs a predetermined signal corresponding to the data. Equipped with a check section that compares with expected values. A voice recognition device test circuit, characterized in that the frequency characteristics and gain characteristics of the voice input section are tested.