JPH11252595A - Voice recognition system having push signal reception function and device realizing the system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce hardware quantity by simultaneously mounting a push signal reception function on a voice recognition hardware and its algorithm and, to handle voice recognition and reception of a push signal as the same function. SOLUTION: Signals from a telephone network are given to a power spectrum conversion section 4, where a power spectrum is calculated, and fed to a push signal detection section 5 and a characteristics amount analysis section 7. A signal detection section 5 discriminates whether or not a received voice signal includes a push signal component. On the other hand, the characteristics amount analysis section 7 extracts the characteristics. A result discrimination section 10 gives a symbol stream of push signals to a result output section 12 when the symbol streams of push signals are stored or when a voice period is detected, a voice recognition 11 is started and the recognition result is fed to a result output section 12.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to voice recognition by telephone and reception of a push signal.

[0002]

2. Description of the Related Art Voice message services by telephone are used. This means that when you call a specific number,
You can listen to automatically sent voice messages such as breaking news, weather forecasts, and prize quizzes. here,
The voice message is in the form of a menu, and when the caller sends menu selection information such as a push signal (also called a dial tone or DTMF), another voice message corresponding to the signal can be heard. .

[0003] Recently, the voice recognition performance is almost at a practical level even for telephone voice, and application to the above-mentioned voice message service is being studied. For example, the code of the push signal for calling up the voice message of “Professional Baseball News” is “2689
# "," Professional baseball "is" 2689 # "
Without this association, the voice message service cannot be used. However, if voice recognition is used, it is possible to construct a system in which a voice message of a target professional baseball news can be immediately heard by simply calling the voice message service center and saying only one word "professional baseball".

However, it is difficult for current speech recognition technology to recognize speech having a very short time length. The most typical example is a digit voice. Names from 0 to 9 to be recognized are "Zero", "Ichi", "Ni", "San", "Yon".
When "go", "roku", "nana", "hachi" and "kyu" are used, "ichi" and "hachi" and "san" and "yon" have similar voice characteristics and are difficult to distinguish. Alternatively, since "Ni" and "Go" are very short voices, a failure that is not recognized even when input is likely to occur. In this regard, when the input of a number is required in a voice band equivalent to a telephone, input by a conventional push signal is advantageous with less errors. Therefore, there is an example in which a push signal receiving function is added to telephone voice recognition hardware. For example, a speech recognition board released by NTT Intelligent Technology Co., Ltd. under the name of "parrot board" is one example.

By the way, in the above-described speech recognition board, generally, a hardware portion for speech recognition and a hardware portion for receiving a push signal have a separated configuration. The reason is that the analysis processing in speech recognition (for example, L
PC analysis method) is not suitable for analyzing signal waveforms having a steep spectral component only in a specific frequency band such as a push signal, and an LSI dedicated to receiving a push signal is commercially available at a low price. by.

Conventionally, however, the hardware for voice recognition is a digital signal processing LSI (DSP).
A push signal receiving LSI is an analog input to enhance versatility even if a DSP is internally used, while a digital signal input is required. That is, a separate waveform conversion circuit and a separate voice branching circuit are required to capture voice from the telephone line into the respective hardware, and the amount of hardware increases as a whole. In particular, the new form of telephone, computer
In the telephony integration (CTI) hardware, since the voice is processed by a digital signal, the push signal receiving L which requires an extra D / A conversion circuit is required.
SI is difficult to mount on CTI hardware.

[0007]

As described above, conventionally, the voice recognition hardware and the push signal receiving hardware each required different signal input means. There was a problem that the wear scale became large. By the way, recently, the speed of the DSP for speech recognition has been increased, and real-time processing has become possible even for processing requiring a certain processing capacity, such as fast Fourier transform. The fast Fourier transform is an example of a calculation method for analyzing a speech spectrum in a small band, and can accurately detect a push signal having a steep spectrum component only in a specific frequency band. Therefore, if the push signal can be analyzed and detected in the speech analysis process for recognition, it is not necessary to add hardware for receiving the push signal, and from the application handling speech recognition, speech recognition and push signal reception are the same. Can be treated as a function.

An object of the present invention is to simultaneously implement a push signal receiving function in voice recognition hardware and an algorithm as described above to reduce the amount of hardware and to perform the same function in voice recognition and push signal reception. It is to be treated as.

[0009]

According to the present invention, the input voice from the voice input means is converted into a digital waveform by the conversion means, and the power spectrum of the digital waveform is obtained by the power spectrum conversion means. , The push signal component is detected by the push signal detecting means, the detected push signal is converted into a symbol string by the push signal storing means and stored, and at the same time, the voice feature is derived from the power spectrum by the feature analyzing means. Then, the voice feature amount is stored in the analysis result storage unit, and which is a valid input result is determined by the determination unit from the detected symbol sequence of the push signal and the detected voice feature amount, and is stored. The voice recognition is performed by the recognition means using the voice feature amount, and the symbol sequence of the push signal or the voice recognition result is output to the output means. It is more output. In the present invention, a speech signal is received from a telephone line, and a short-time spectrum analysis is performed as a speech analysis technique for recognition. If a push signal component is detected from the spectrum analysis result, conversion and accumulation of the push signal into symbols are performed, and if a voice feature is detected without observing the push signal, recognition processing is performed on the detected voice section. .

[0010]

FIG. 1 shows a voice recognition apparatus according to an embodiment of the present invention. A line input terminal 1 is a terminal connected to a telephone line 22 via a public network 21, and a line interface unit 2 performs call control on the connected telephone line 22 and supplies telephone voice to a voice waveform conversion unit 3. The voice waveform conversion unit 3 converts the telephone voice into a digital waveform that is easy to handle by signal processing in the subsequent stage,
Divides the audio waveform converted in the previous stage into short time sections,
The power spectrum waveform is obtained and supplied to the push signal detection unit 5 and the feature value analysis unit 7. Push signal detector 5
Detects a push signal component present in the power spectrum waveform. The push signal storage unit 6 stores the push signal detected digit by digit as a digit string. The feature analysis unit 7 extracts a speech feature required for recognition from the power spectrum waveform, and supplies the speech feature to the speech section detection unit 8 and the analysis result storage unit 9. The voice section detection section 8 specifies a voice section required for recognition, and stores the extracted voice feature amount in the analysis result storage section 9 for the detection section. The result determination unit 10 determines whether the voice feature value or the push signal is valid for the voice analysis result, and if the voice feature value is determined to be valid, the voice recognition unit 11 executes voice recognition. The output unit 12 outputs a speech recognition result or a push signal reception result.

The operation of the apparatus shown in FIG. 1 will be described below. When a caller makes a call to the voice recognition device 23 of the present invention from the telephone 24, the communication call is connected to the line interface unit 2 via the line input terminal 1, and the voice communication path is connected to the voice waveform conversion unit 3. You. The voice of the caller or a push signal generated by the caller's operation is converted into a digital waveform by the voice waveform converter 3. If the telephone line input to the line input terminal 1 is an analog network,
The operation of the audio waveform converter 3 corresponds to A / D conversion. When the telephone line is an ISDN network, the operation of the audio waveform converter 3 is a conversion operation from a nonlinear digital waveform to a linear digital waveform. Here, the sampling frequency of the digital waveform is, for example, 8 kHz. Audio waveform converter 3
The output digital waveform is divided into superimposed sections (referred to as frames) having a length of 32 msec at intervals of, for example, 8 msec, and is input to the power spectrum calculation unit 4 to obtain power spectrum information of voice. The method of calculating by the power spectrum conversion unit 4 is, for example, a spectrum calculation method well known as fast Fourier transform, and for example, an autocorrelation function which is a feature amount physically equivalent to the power spectrum. The result of this calculation is sent to the push signal detection unit 5, and it is determined whether or not the input audio signal contains a push signal component. The push signal has a high group F1 [Hz] and a low group F2 shown in FIG.
It is determined as a continuous waveform combining pure tones of two frequencies of [Hz]. It is determined that a continuous signal having components of F1 and F2 has a constant power and lasts for a predetermined time or more. , F2, even if it is a continuous signal, the signal (for example, real voice) having a power equal to or higher than a certain value in other bands is not determined. Even if the power of F1 or F2 changes in a very short time, it is different. A condition is required to avoid erroneous determination (typically a phenomenon called chattering) when the push signal is input. FIG. 3 shows an example of a flowchart of a specific determination algorithm. The push signal detector 5 first receives the short-time power spectrum (Step 1) and observes the signal power Pow in the short-time section (Step 2). For the combination of the high group frequency F1 and the low group frequency F2 shown in FIG.
Is estimated, and the maximum power M obtained by adding the signal powers of only the F1 and F2 bands at that time is derived (step 3). Further, the maximum power Q in the frequency band from 300 Hz to 600 Hz is observed (step 4). This band is a band between the lower limit of the telephone voice band (300 Hz to 3.4 kHz) and the lower limit of the push signal frequency F1 in FIG. Q is a value of M or more when a voice is observed instead of a push signal. Therefore, the signal power Pow is equal to or greater than the threshold value A, and the power ratio M / Q is equal to the threshold value B.
It is determined that the above condition (B> 1) is set as the push signal receiving condition (step 5).

Thereafter, it is determined whether the received push signal is the same as the push signal observed immediately before (step 6). If the combination of F1 and F2 is different, the information is input as a new digit and the push signal is used. It is transmitted to the storage unit 6 (step 7), and if the same, the same push signal is continuously transmitted (that is, a state where the finger is kept pressed without leaving a certain push button). Consider and ignore. If the signal is not received as a push signal in step 5, it is determined whether the signal power M is lower than the lower threshold C (step 8). If it is lower, the input of the push signal is not performed at all (that is, the finger is released from the push button), and the previously input push signal information is discarded (step 9). And ignore it. This has the effect of preventing chatter as described above. When these determination processes are completed, the push signal detection unit 5 returns to a state of waiting for the signal power of the next short frame (step 1).
0).

According to the above-mentioned algorithm, the input of a new digit of the push signal is stored in the push signal storage unit 6 as the symbol string shown in FIG. Here, a symbol corresponding to the end of the input is reserved in the push signal storage unit 6 (referred to as a delimiter). When the delimiter is transmitted from the push signal detection unit 5, the symbol sequence of the push signal stored up to the delimiter is stored. The result is sent to the result determination unit 10. At the same time, the push signal storage unit 6 observes the time (interdigit time) from the last input of the push signal to the present time. When the interdigit time reaches a predetermined value, the input of the push signal is determined to be completed. And sends the symbol string of the push signal stored in the result determination unit 10.

On the other hand, the calculation result of the power spectrum conversion unit 4 is branched and input to the feature value analysis unit 7 to extract features for speech recognition. The feature extraction method is, for example, a method well known as log power or, for example, LPC cepstrum. At this time, a feature amount that easily reflects voice section information, such as logarithmic power, is input to the voice section detection unit 8 and used for detecting section information. The section detection method is, for example, a method in which all sections having power equal to or higher than a certain value are regarded as speech. For example, the voice section detection unit 8 may detect only the beginning of the voice depending on the algorithm of the voice recognition unit 11 and consider a certain time length thereafter as a voice section. For the section determined to be speech by the speech section detection unit 8, the speech feature amount output from the feature amount analysis unit 7 is stored in the analysis result storage unit 9.

The result determination section 10 receives the push signal accumulation information of the push signal accumulation section 6 and the section detection information of the voice section detection section 8, respectively, and judges which result is a valid output result. Since a push signal has a certain signal power and a certain duration like voice, if the push signal is accumulated, it is highly likely that the push signal is erroneously determined as a voice section. Therefore, it is more appropriate for the result output to give priority to the push signal output.
A value of 0 sends the symbol sequence of the push signal to the result output unit 12 if the symbol sequence of the push signal is stored, and activates the speech recognition unit 11 if the symbol sequence is not stored. In addition, the result determination unit 10 can output a result of only voice recognition or only a push signal. The recognition method used in the voice recognition unit 11 is, for example, a template matching method, and is, for example, a statistical matching method called HMM. In general, the voice recognition process has a larger calculation scale than the voice analysis process, so if it is determined in advance whether it is a voice signal or a push signal, unnecessary calculations can be avoided. The result of the speech recognition is sent to the result output unit 12, and the result output unit 12 sends the vocabulary name and the symbol string of the push signal, which are the result of the speech recognition, to an external device requiring the speech recognition device 23.

As shown in FIG. 1 as a control unit 14 in the speech recognition apparatus of the present invention, the control unit 14 is mainly composed of a microprocessor, for example, and controls each unit sequentially. It performs reading and writing of various storage units, that is, this device can be operated by decoding and executing programs by a computer.

[0017]

As described above, according to the present invention, (1) the hardware required for receiving a push signal is included in the hardware for speech recognition as compared with the conventional speech recognition apparatus, and the amount of hardware is prevented from increasing. (2) Since the voice recognition algorithm includes the push signal receiving algorithm, the application can call the voice recognition and the push signal reception in a common procedure, and can use only one or both of them as information input means. have.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a functional configuration of an embodiment of a speech recognition device according to the present invention.

FIG. 2 is a diagram showing a relationship between a symbol of a PB signal and its frequency.

FIG. 3 is a flowchart showing an example of an operation procedure of a push signal detection unit 5 and a storage unit 6 in FIG.

Claims (4)

[Claims] 1. A voice input means, a conversion means for converting an input voice from the voice input means into a digital waveform, a power spectrum conversion means for obtaining a power spectrum of the digital waveform, and a push signal component from the power spectrum. Push signal detecting means for detecting, push signal accumulating means for converting the detected push signal into a symbol string and accumulating the same, characteristic amount analyzing means for deriving an audio characteristic amount from a power spectrum, and accumulating the audio characteristic amount Analysis result accumulating means, and a judgment means for judging which is a valid input result from the detected symbol sequence of the push signal and the detected speech feature quantity, and a speech feature quantity of the analysis result accumulation means. Recognition means for recognizing the voice signal by using the symbol sequence of the push signal according to the determination result of the determination means. Or an output means for outputting the result of voice recognition, wherein the voice recognition device has a push signal receiving function. 2. A process for obtaining a power spectrum from an input digital waveform, a process for detecting a push signal component from the power spectrum, a process for converting the push signal into a symbol string, and storing the symbol sequence in a push signal storage means. A feature value analyzing step of deriving a voice feature value from the power spectrum; a step of storing the voice feature value in an analysis result storage unit; a symbol sequence of the detected push signal and the detected voice feature value A determination step of determining which is a valid input result from the above; a step of outputting the symbol string of the push signal when the determination determines that the symbol string of the push signal is valid; If it is determined that the volume is valid, voice recognition is performed using the voice feature amount stored in the analysis result storage unit, and the result is determined. Recording medium for recording a program for executing a recognition process of outputting, with computer. 3. The recording medium according to claim 2, wherein the program includes executing the recognition step when it is determined that the signal is not a push signal. 4. A voice section is detected from the voice feature amount,
5. The recording medium according to claim 2, wherein the program includes executing a voice section detection step of using the detected voice section information for determining validity from the voice feature amount in the determination step.