JPS59121398A - Voice recognition system - 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 (a) Technical Field of the Invention The present invention relates to a speech recognition device, and particularly to a speech recognition method capable of identifying each speaker among a plurality of speakers.

b. Prior Art and Problems There are two types of speech recognition methods: a speaker-independent method and a specific speaker method, in which the recognition device performs speech recognition by comparing the input speech with a speech pattern dictionary prepared in advance.

Using the same speech pattern dictionary as the speaker-independent method,
This method recognizes the voices of unspecified speakers. Although this method can recognize the voices of an unspecified number of speakers, it is not possible to obtain a high recognition rate because the speech pattern dictionary is common.

The specific speaker method is a method in which a speech Bakun dictionary is registered for each speaker, and speech is recognized using a speech pattern dictionary corresponding to the speaker. Although this method requires the work of registering a speech pattern dictionary for each speaker, it is possible to obtain a high recognition rate.

By the way, one application field of speech recognition devices is the task of creating minutes of meetings and the like. In this way, when multiple speakers are present and talking, it is necessary to identify which speaker's voice is being input into the speech recognition device before performing speech recognition. be. Therefore, in such a case, the following voice recognition method can be considered.

(1) Recognize each person's voice using a speaker-independent method.

(2) For each speaker, prepare a speech pattern dictionary and a recognition device corresponding to that speaker, and recognize each person's speech using a speaker-specific method.

(3) For each speaker, prepare a speech pattern corresponding to that speaker and one recognition device, and before each person inputs speech, use which speech Bakun dictionary to perform speech recognition. While inputting personal identification information indicating the identity of the user, a voice pattern dictionary is specified, and each person's voice is recorded using the specified voice pattern dictionary.

However, (1), (2), and (3) each have the following drawbacks. In (1), the number of words that can be recognized is small and the recognition rate is low. In (2), the speech pattern dictionary,
This requires as many recognition devices as there are speakers, which poses a cost problem. In (3), inputting personal identification information is cumbersome;
Additionally, users may forget to input personal identification information before inputting voice.

C.Object of the Invention Accordingly, in the present invention, the above-mentioned drawbacks can be overcome. This paper proposes a speech recognition method that can identify each speaker among multiple speakers.

d.Structure of the Invention Therefore, in the present invention, the input means selects one speech pattern dictionary from a plurality of prepared speech pattern dictionaries, and the recognition means selects one speech pattern dictionary from among a plurality of speech pattern dictionaries, and uses the speech pattern dictionary and the speech inputted by the input means. In a speech recognition device that performs speech recognition without comparison, a plurality of speech input means are provided, and a speech pattern is determined based on the phase difference or the ratio of input levels of the same speech input multiple times from at least two of the speech input means. We propose a speech recognition method that is characterized by dictionary selection.

e Embodiment 1 of the invention Section 1 shows a speech recognition device using a speech recognition method which is an embodiment of the invention, and 1 and 2 are my own.
□, 3I/'i speaker determination section, 4 is a feature extraction section, 5 is an authentication section, 6 is a registration section, 7 is a control section, 8 is a keyboard display, 9 is a dictionary section, IO is a switching section, 11 13 indicates a speech pattern dictionary, and 14 indicates a speech recognition device.

Figure 2 shows the positions of speakers and microphones when implementing this embodiment, where A, B, and C are speakers, l and 2 are microphones,
14 voice recognition devices are shown respectively.

FIG. 3 shows the waveforms of voices input by speakers A, B, and C from microphones 1 and 2, and FIG. 4 shows the waveforms of voices input by speaker A from microphones 1 and 2. -1 is the waveform of the voice input by 9 speakers A with more than microphone 1, A-2 is the waveform of the voice input by 2 speakers A with 25 microphones, B11d is the waveform of the voice input by speaker B from microphone 1, B-2 is the waveform of the voice that speaker B inputs more than microphone 2, C-1u, the waveform of voice that speaker C inputs more than microphone 1, and C-2 is the waveform of voice that speaker C inputs more than microphone 2. Each wave of input audio is shown.

In addition, FIG. 4 is similar to FIG. 3 A-1. An enlarged view of A-2 is shown.

In the present invention, two microphones are used as voice input means. Then, the speakers and twelve microphones are arranged so that the distances 4 between the plurality of speakers (three people) and the two microphones are different. Then, a difference occurs in the phase or input level of the same voice d1 of each speaker inputted by the two microphones. The phase or magnitude relationship of the input level is made to correspond to personal identification information.

As can be seen from the positions of speaker A, B, and C microphones 1 and 2 in Figure 2, if we compare the distances between speaker A and microphones 1 and 2, we find that: Close to A. Therefore,
Figure 3 A-1. As shown in A-2, if the point in time when the voice waveform exceeds a certain level (determined by considering the surrounding noise level) is the point in time when the voice starts (starting point), the input from microphone 1 The starting point of the audio waveform is
The starting point of the waveform of the voice input from the microphone 2 is also earlier in time. Here, you can see 51C by looking at the crocodile diagram 4.
There is a difference in the input level between A-1 and A-2 (the input level of A-2 is smaller), but if the starting point level is the same, the time difference will be emphasized, so the input level will be adjusted according to the input level. There is no need to change the level of the starting point.

Similarly, as shown in B-1 and B-2, for speaker B's voice, the starting points ■ and ■' are almost simultaneous, and C-1°C'-
As shown in 2, for speaker C's voice, the starting points ■ and ■
′ is slower than ■ than ■′.

Therefore, the microphone 1 as mentioned above. The phase difference of the input voice of less than 2 is used in correspondence with personal identification information.

For speakers A, B, and C, before starting speech recognition, the voices are registered in dictionaries 11 to 13. For registration, the keyboard 8 instructs the switching unit 10 to select one of the dictionaries 11 to 13. Each speaker has one microphone.
Or input voice (50 sounds, numbers, etc.) using microphone 2.

The extraction unit 4 extracts the frequency characteristics of the input voice. The extracted frequency characteristics are registered in the selected dictionary by the registration unit 6. Dictionaries 11-13 are
A has a dictionary 11. B has a dictionary”2. It is assumed that the C language dictionary 13 corresponds to this. When speakers A, B, , , and C take positions as shown in Figure 2, before starting a conversation, A, B, and C make some sounds (for example, pronouncing their names) through microphones 1 and 2, respectively. Enter a small amount. The determining unit 3 detects the phase difference between the voices input from the microphones 1 and 2 for the voices A, B, and C, respectively. At the same time, enter the personal identification information on the keyboard 8, and enter the personal identification information (kana, numbers, etc.) in the judgment section 3.
and the phase difference are stored in correspondence. Further, this personal identification information corresponds to dictionaries 11, 12, and 13, respectively. A,
When B and C start a conversation, the determining unit 3 determines the speaker based on the phase difference between the voices input from the microphones 1 and 2. The determination unit 3 sends personal identification information to the dictionary unit 9. In the switching unit 9, a switching unit 10 selects one based on the personal identification information.
Select one dictionary from dictionaries 1 to 13. In the recognition unit 5, the frequency characteristics of the voice extracted by the extraction unit 4 are
The audio is recognized while being compared with the selected dictionary. Here, the audio input to the extraction unit 4 is the audio with a faster phase among the audio inputs from the two microphones, and if there is no phase difference,
A small amount of audio is input from a predetermined microphone. or,
The determining unit 3 instructs the extracting unit 4 to extract features from the audio input from the two microphones.

As another embodiment of the present invention, there is a method for comparing the input levels of voices input from two microphones. That is, as the distance between the speaker and the microphone increases, the input level decreases, and as the distance decreases, the input level increases. Using this relationship, the input levels of the same voice input to microphone 1 and microphone 2 are compared, and the speaker is identified based on the magnitude relationship. Looking at Figure 3, for speaker A, the human power level of microphone 1 > the input level of microphone 2, for speaker B, the input level of microphone 2 among the input levels of microphone 1, and for speaker C, the input level of microphone 1 The following relationship holds true: level <input level of microphone 2.

Therefore, the magnitude relationship of these input levels is described first.
Similar to the phase difference relationship in the fc embodiment, it is made to correspond to personal identification information.

The method of measuring the ratio of input levels is described in FIG.

First, the point in time when the waveform of a voice exceeds a certain level (taking into account the level of surrounding noise) is defined as the point in time at which the voice begins (starting point). Then, at the starting point ■, ■′ (within a certain period of time) determined for the audio input from microphone 1.2,
Find the peak values (Ll, Lt) of both inputs. Then, the level ratio of LH and L2 is determined, and the ratio is associated with personal identification information.

In addition, when identifying speakers using the difference in input levels in this way, the voice input to the extraction unit 4- is the voice input with a higher input level, and if the input level power is the same, , the audio input is from a predetermined microphone.

In this embodiment, a case where there are three speakers is shown, but the same implementation is possible even when there are more than three speakers by dividing the phase difference and the input level ratio more finely.

In this embodiment, two microphones are used, but if three or more microphones are used, more information can be used to identify the speaker.

e Effect of invention According to the present invention, large costs may be required.

[Brief explanation of the drawing]

FIG. 1 shows a speech recognition device used in a speech recognition system that is an embodiment of the present invention, 1 and 2 are microphones, 3 is a speaker determination section, 4 is a feature extraction section, 5 is a recognition section, and 6 is a Registration section, 7 is control section, 8 is keyboard display 1.9 is dictionary section, 1
0 represents a switching unit, 11 to 13 represent voice pattern dictionaries, and 14 represents a voice recognition device. FIG. 2 shows the positions of the speaker and the microphone when implementing this embodiment.1. A, B, and C are speakers, 1 and 2 are microphones, and 14 is a speech recognition device, respectively.

Claims (1)

[Claims] The input means selects one voice butter dictionary from a plurality of prepared voice pattern dictionaries, and performs voice recognition while the recognition means compares the voice pattern dictionary with the voice. phase difference of the same voice, or
A speech recognition method characterized by selecting a speech pattern dictionary based on the ratio of input levels.