JPS607498A - Word voice recognition equipment and method - 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] Technical Field of the Invention The present invention relates to a speech recognition device, and more particularly to a speech recognition device for inputting and recognizing words by speech. The present invention relates to a speech recognition device that does not require registration of all recognition target words in advance.

TECHNICAL BACKGROUND, PRIOR ART, AND PROBLEMS Conventional word speech recognition devices generally include all-word registration type speaker specific recognition devices and non-specific speaker recognition devices. The former method employs a method in which a speaker is identified and all words to be recognized are input and registered in advance. This type of device has a good recognition rate and can set a large number of recognition target words, but since it is necessary to register the voices for all recognition target words in advance, it is difficult for users to recognize recognition target words that exceed 100 words. When users register their voices, there is a problem in that the burden of registration on users is too large.

On the other hand, the latter has the advantage that the speaker is not specified and the user does not need to register the voice in advance. This method has the problem that the number of words to be recognized is limited to about 10, and the problem that the recognition target words cannot be changed freely.

OBJECTS OF THE INVENTION An object of the present invention is to solve the above-mentioned problems of the prior art and to provide a word speech recognition device that can recognize a large number of words with a minimum number of speech registrations. Another object of the present invention is to provide a speech recognition device that can easily change the recognition target word.

The above object of the present invention is to perform segmentation of phonemes at the time of registration, without adopting a method of performing phoneme segmentation during speech recognition as in conventional methods, and to create a phoneme dictionary in which segmentation has been performed in this way. This is achieved based on the idea of recognizing word audio input using .

Structure of the Invention In the present invention, a plurality of channels of audio data signals are obtained by receiving an audio input and rectifying the audio input through a plurality of parallel bandpass filters having different center frequencies, and a plurality of audio data signals from the plurality of audio data signals. A voice input means for generating one nof word signal, applies a voice input for a registered word that is independent of the word to be recognized to the voice input section, and generates a z4 word signal of the voice input for the registered word. a labeling means for creating a speech dictionary from a plurality of speech data signals of speech inputs for registered words based on temporal changes; A phoneme lattice that creates phoneme data for the voice input of a recognized word from a power signal as an output and a plurality of voice data signals, and creates phoneme lattice data for phonemes with a high degree of approximation to phonemes in the phoneme dictionary obtained in advance. A word speech recognition device is provided that includes a creation means and a label matching means for matching the phoneme lattice data with pre-stored recognition target words and extracting all words with the highest degree of similarity.

In addition, in the present invention, there is a step of voice inputting registered words independent of the words to be recognized and creating a phoneme dictionary based on the voice input, and a step of voice inputting the words to be recognized and creating phoneme data based on the voice input. There is also provided a word speech recognition method comprising the steps of creating a phoneme lattice by comparing the phoneme lattice with the phoneme dictionary, and comparing and comparing the phoneme lattice with the recognition target word to extract a word with the highest degree of similarity.

Embodiment of the Invention An embodiment of the present invention will be described below with reference to the accompanying drawings.

FIG. 1 shows a block diagram of a speech recognition device according to the present invention. The speech recognition device includes a microphone 1 for inputting speech, a speech input section 2 that receives a signal from the microphone and processes the input signal, and an output signal S2 of the speech input section 2 and a signal from a registered word storage section 8. It has a labeling section 5 that creates a phoneme dictionary based on S8 and stores it in a phoneme dictionary storage section 6. Furthermore, when there is a speech input to be recognized, the speech recognition device includes a phoneme lattice creation section 3 that creates a phoneme lattice by referring to the phoneme dictionary data of the phoneme dictionary storage section 6 from the signal S2 from the speech input section 2; A label matching unit 4 that collates the phoneme lattice with the recognition target word in the recognition target word storage unit 7 to create a corresponding word, and a display of the recognized word or a word to be registered, or between it and a low user. A display 9 for man-machine communication, the key having a key (not shown).

FIG. 2 shows an internal circuit diagram of the audio input section 2. In FIG. 2, filter channel circuits 211 to 21n and the circuit 2
An adder 22 is shown which sums the outputs of 11 to 21n. Further, the details of the circuit 211 among the circuits 211 to 21n are shown in FIG. In the third diagram, BPF
I is a bandpass filter with center frequency f1, D is a diode, and R
Vi resistor, C indicates capacitor or capacitor. Other circuit 212
~2in is also similar to the circuit configuration shown in Fig. 3, but the center frequency of the IW band filter is different from /1 and is respectively f2~/n.
'''C is there.

As is clear from the circuits in FIGS. 2 and 3, when the audio input (AC) from the microphone 1 is input to the audio input section 2, it is bandpass filtered for each channel, rectified by the diode D, and the DC are smoothed by an RC time constant circuit, and further converted into digital signals through digital lines 231-23n to become output data signals SDI-SDn of each channel. On the other hand, when the signals SDI to SDn are applied to the adder 22, its output represents the power (S
PwR). The S crystal claw is also converted into a digital member (im number) via the digitizer 24.The output signal S2 of the audio input section 2 is the signal S D1 to SDn%5PWR.
This is a general term for the following.

The operation of the apparatus shown in FIG. 1 will be explained below. The device R:
The form in which speech recognition is performed using is broadly divided into a registration phase and a recognition phase.

First, the operation of the registered 7Aze will be described. The user (speaker) is identified, and the following steps are performed for the identified user. In the registration phase, phoneme segmentation, which will be described later, is performed to create a phoneme dictionary, but speech registration words that are not needed for phoneme dictionary creation are determined in advance and stored in the registration word storage section 8.

A voice registered word, for example, ``r AGA'' from the registered word storage section 8 is displayed on the love display 9 and the user is instructed to speak it. The user speaks the voice registration K into the microphone 1. The audio input signal Sl of the microphone l is applied to the audio input section 2, and the respective channel data S
DI to SDn and 4W-8PWR are obtained. These signals are applied as 82 to the labeling section 5. /? The mulberry PWR is shown as a characteristic curve as shown in FIG. 4 (4). In FIG. 4 (4), the horizontal axis t represents time.

In the labeling section 5, a constant time interval Δt with respect to time t
Then, the change dSPWR'(z) of adjacent equally spaced frames called child frames is determined (FIG. 4(B)).

Based on the thus-obtained /yawa change d S PWR and the symbol string of the word rAGAJ, a label is attached to the phoneme of each frame, rAJrGJ in the example shown in FIG. 4(C). Here, labeling is performed so that the peak value of the power change is consistent with the symbol string of the word. After phoneme labeling is done in this way (Figure 4 (C))
, each frame is stored in the phoneme dictionary storage unit 6 as a phoneme dictionary in the form shown in FIG.

The phoneme dictionary shown in FIG. 5 will be described in detail. In FIG. 5, rSE LBLJ represents a phoneme label, and the phonemes at this time include a vowel "A" and a prefix "G" such as a voiced consonant "Gagig" for each frame. Also rs
E DATJ indicates phoneme data, and for one phoneme rAJ that has changed by 1,000 times in the frame, it is set as corresponding data 1DAT1 among the channel data SDI to SDn, and is data representing "characteristics."

Once a phoneme dictionary is created for "AGA" as described above, a phoneme dictionary is created for the next registered word in the same manner. Here, the voice registration words are not equivalent to the target words to be recognized, but rather have an independent relationship, and may be as many as necessary to create a phoneme dictionary necessary for recognizing all target words. In other words, even if the number of voice registration words is sufficient to create a certain amount of phoneme dictionary, and the number of words to be recognized increases, there is no need to perform further voice registration.

In addition, at the time of creating the phoneme dictionary as described above, if the labeling is inconsistent with the symbol string for the registered phonetic word known in advance, the user is prompted to repeat it via the display 9. , the voice registration process is returned to ensure the creation of a phoneme dictionary.

Next, the recognition phase in which speech recognition is actually performed using the phoneme word t will be described.

Assume that a voice input to be recognized is applied to the microphone 1, and the output signal Sl is obtained as shown in FIG. Output signal S2 of audio input section 2, 5 in FIG.
pNvR2 is applied to the phoneme lattice creation section 3. In the phoneme lattice creation section 3, all phoneme data are generated for each frame (Δt) for the power signal shown in FIG. 6, and distance calculations are performed using a phoneme dictionary obtained in advance. Then, a phoneme lattice table organized in descending order of distance between phonemes (
Figure 7 1) is created. In Figure 7, the horizontal direction is time;
That is, the number of frames and the phonemes arranged in ascending order of distance d in the vertical direction and their distances m are shown. For example, for frame 1 in the leftmost column of FIG. 7, the phoneme "O".

It is shown that a phoneme lattice with distance = 50, phoneme "A", distance = 52, phoneme roJ distance = 53, phoneme "A", distance = 70 has been obtained.

The phoneme lattice table obtained in this manner is compared with the character string of the target word stored in the recognition target word storage unit 7, and the comparison is performed in the all-label collation unit 4. The recognition target word storage unit 7 stores all words to be recognized. In addition, each of these words has a matching window W as shown in FIG.
INDW is provided. Figure 8 shows an example word f-
Regarding 00SAKAJ, each phoneme rOJ, r
OJ, rsJ, rAJ.

A frame band to be compared and verified for each rKJ and rAJ is provided, and verification is performed in an appropriate frame range (the shaded area is outside the range of comparison and verification) to contribute to improving the recognition rate and increase efficiency.

For the phoneme lattice element (Fig. 10) consisting of
Let the minimum distance of the phoneme label that matches the character of the string of KAJ be the distance of that frame, and the distance (
d) and the phoneme label (SE, LBL)'i word (see Figure 9). Next, the sum of these distances is taken, and the value Σd l /n, which is divided by the number of frames n, is set as the distance between the word character string to be recognized and the phoneme lattice formed by the input speech.

The above distance calculation is performed for all word strings to be recognized, and the one with the smallest distance is extracted as the one with the highest degree of approximation and displayed on the display 9 as the recognized word S4.

In this example, roO8AI(A(Osaka)" is the recognized word.

If the recognition word is desired, the user is instructed to that effect by pressing the key or the like, and if not, a retry is made.

Effects of the Invention As described above, according to the present invention, instead of registering the voices of the words to be recognized, it is only necessary to register the necessary number of voices of the words to be registered to create a phoneme dictionary. Since it can be used as is even if the number of target words increases, it is possible to recognize a large number of single and complete words by registering appropriate speech.

Further, according to the present invention, the recognition target words are not registered as voices, but the recognition target words are independent, and the recognition target words can be easily increased, decreased, or changed.

[Brief explanation of drawings]

Figure 1 shows a speech recognition device (4) as an embodiment of the present invention.
Figure 2 is a circuit diagram of the audio input section in the first mounting unit 7, Figure 3 is a detailed circuit diagram as an example of the circuit in Figure 2.2, and Figure 4 is a circuit diagram of the circuit in Figure 2. Figure 4 (B) shows the power characteristic diagram for each frame as the output of Figure 4 (4).
Arm waa change characteristic diagram, 4th ul (c) is a diagram showing the phoneme data obtained based on 4th ul (B), and 5th is an example of a phoneme dictionary based on 4th ul (C). Figure 6 is a power characteristic diagram of the voice input to be recognized, Figure 7 is a diagram showing the Vi element table obtained from the data shown in Figure 6, and Figure 8 is a diagram showing the power characteristics of the voice input to be recognized. A diagram showing an example. Figure 9 is a diagram showing the phoneme data obtained by applying the matching window of Figure 8 to the phoneme lattice data shown in Figure 10. Figure 10 is a phoneme lattice for the speech input to be recognized. It is a figure showing a table. (Explanation of symbols) 1...Microphone, 2...Audio input section, 3...
- Phoneme lattice creation section, 4...Label matching section, 5...
- Labeling section, 6... Phoneme dictionary storage section, 7... Recognition target word storage section, 8... Registered word storage section, 9... Distle-. Patent applicant Fujitsu Ltd. Patent attorney η' Akira Ki Patent attorney Kazuyuki Nishidate Patent attorney 1) Yukio Patent attorney Akira Yamaguchi Figure 4 (A) Figure 5 Ratio Figure 6

Claims (1)

[Claims] 1. All audio inputs are input, and the audio inputs are rectified through a plurality of parallel bandpass filters with different center frequencies, and the audio data of the plurality of channels are rectified. audio input means for generating one noise signal from the signal;
A voice input for a registered word that is in an independent relationship with the word to be recognized is applied to the voice input section, and the voice input for the registered word is determined based on the temporal change in the voice input signal for the registered word. Labeling means for creating a phoneme dictionary from a plurality of audio data signals; A phoneme lattice creation means for creating phoneme data regarding voice input; creating phoneme lattice data for phonemes with a high degree of similarity to phonemes in the phoneme dictionary obtained in advance; and recognition in which the phoneme lattice data is stored in advance. A word speech recognition device that is fully equipped with a label matching means that matches the target word and extracts the word with the highest degree of similarity. 2. The step of voice inputting a registered word independent of the word to be recognized and creating a phoneme dictionary in advance based on the voice input; inputting the word to be recognized voice and creating a phoneme r-y based on the voice input; A word speech recognition method comprising the steps of creating a phoneme lattice by comparing it with a phoneme dictionary, and comparing and comparing the phoneme lattice with a recognition target word to extract a word with the highest degree of similarity.