JPS59119397A - 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] Technical Field of the Invention The present invention is a speech recognition device, in particular, the entire device is divided into an analysis section, a recognition section, and an input/output management section, and the processing results are configured to be inherited from one level to the next. The present invention also relates to a speech recognition device that speeds up transfer between an analysis section and a recognition section.

(B) Technical Background and Problems Traditionally, speech recognition devices generally perform frequency analysis of input speech to generate multiple pieces of time-series feature information and store them in a (++) dictionary. The voice recognition device is configured to determine the category of the input voice by matching it with the standard feature time series that has been provided, and output the recognition result to the input/output device.In such a voice recognition device, Each of the above processes (+) (ii) (ii[
), a configuration is being considered in which a microprocessor is prepared for each process and the results of each process are sequentially taken over.

When considering such a configuration, the problem is how to take over the processing results as one unit in each process. That is, the above processing (i) (ii) (iiD
With reference to FIG. 1, problems in handing over processing results in the case of an analysis section, a recognition section, and an input/output management section will be explained with reference to FIG.

In FIG. 1, it is assumed that the input voice is something like a figure. It should be noted that the illustrated waveform may be considered to be a case of an input voice in which a plucked sound such as "Ataru" is mixed, for example. As is well known in the art, the analysis section performs frequency analysis of the input voice and generates time-series feature information for each frequency band, but also monitors, for example, the energy of the input voice.
The length of the silent period below the indicated threshold is checked. For example, if a silent period exists for 0.3 seconds or more, it is considered to be a silent period between words, and if it is shorter than that, it is assumed that it is caused by the above-mentioned clipped sound, and it is determined that the word is still continuing. It will be done like this.

Since the silent period (5) shown in FIG. 1 is within 0.3 seconds, the analysis unit considers that the word continues. Then, since the first 0.3 seconds in the silent period (B) in the diagram is detected, it is assumed that a word break has arrived, and the time series feature information of each frequency band up to that point is divided into 16, for example, and 16 steps of information are generated. The information is passed to the recognition unit as follows. In the case shown in FIG. 1, the recognition unit starts receiving the information of the 16 steps after the lapse of 0.3 seconds. The recognition result is then delivered to the input/output management section.

As described above, when attempting to adopt a method of sequentially handing over processing results between the analysis section, the recognition section, and the input/output management section, the handover mode as shown in FIG. 1 is generally inevitable. However, when trying to perform online speech recognition processing, the delay time of 0.3 seconds mentioned above also becomes a problem.

(6) Object and Structure of the Invention The present invention aims to solve the above-mentioned problems.The speech recognition device of the present invention performs frequency analysis of input speech to obtain a plurality of time-series feature information. In the speech recognition device configured to generate the above-mentioned input speech, determine the category of the input speech by matching it with the standard feature time series stored in the dictionary, and output the recognition result to the input/output device. , an analysis section that generates the plurality of time series feature information, a recognition section that determines the category, and an input/output management section that outputs the above, and the results obtained by the analysis section are sent to the recognition section. The analysis section monitors the length of the silent period of the input voice. comprising a visual means, and transmitting the time-series feature information based on the processing results up to that point to the recognition unit at the start of the silent period;
The recognition unit is configured to notify the recognition unit whether or not the silent period is an inter-word silent period, and if it is not an inter-word silent period, transfer the time-series feature information again in the next silent period, and the recognition unit The present invention is characterized in that the recognition process is started based on the received time-series feature information, and it is determined whether to continue or stop the recognition process at the time of notification corresponding to the inter-word silent period. This will be explained below with reference to the drawings.

(2) Embodiment of the Invention FIG. 2 shows the configuration of an embodiment of the present invention, and FIG. 3 shows an explanatory diagram illustrating processing of an embodiment of the present invention corresponding to the explanatory diagram shown in FIG. 1.

In FIG. 2, 1 is an analysis section, 2 is a recognition section, 3 is an input/output management section, 4 is an Iri network node that mediates transfer between each section, and 5 is an analysis circuit that supports multiple frequencies. 6 is an analysis processor that performs processing and control in the analysis unit IK; 7 is a silent period monitoring means; 8 is a recognition processor that performs processing and control related to recognition processing. 9 is a dictionary and matching circuit; 10 is an input/output processor which performs processing and control such as outputting characters based on recognition results; and 11 an input/output device.

The operation will be explained below with reference to FIG.

While the analysis circuit 5 is extracting the above-mentioned pieces of time-series characteristic information υ, the analysis processor 6 uses the monitoring means 7 to check for the existence of a silent period. Then, when the illustrated silent period (5) starts, the analysis frame 0
Rosetsuza 6 assumes that the inter-word silent period has started at that point, and divides the time-series feature information at that point into information of 16 steps and 2 degrees as described above (Illustrated analysis data 1
), the recognition unit 2 transfers the recognition 70 to the processor 8. The recognition unit 8 immediately starts the matching process and continues the process. Also, if the processing is completed early, the result is transferred to the input/output processor 10 of the input/output management section 3. During this time,
Assuming that the silent period of the input voice disappears within 0.3 seconds as shown in period (4), the analysis processor 6 issues a cancel command to the recognition processor 8 and, if necessary, to the input/output processor 10. And recognition processor 8
In the input/output processor IO, the processing up to that point is discarded.

Next, when a silent period starts as shown in the illustrated period (B), the analysis processor 6 divides the information from the illustrated analysis start point t1 into information of 16 steps (illustrated analysis data 2) in the same way as above, and the recognition unit 2 It is transferred to the recognition processor 8 of. The recognition processor 8 begins immediate processing. Illustrated silent period (
If the silent period continues for O.3 seconds or more as in B), the analysis processor 6 issues a continuation command to the recognition processor 8 and, if necessary, to the input/output processor 10. Therefore, the process continues to be executed. Of course, after the illustrated silent period of 3 seconds has elapsed, the voice corresponding to the next word may be input.

(g) As described in detail of the invention, 1. According to the present invention, a voice recognition device,
By dividing the system into an analysis section, a recognition section, and an input/output management section, it becomes possible to efficiently transfer information between each section, and it becomes possible to eliminate time delays during online processing.

[Brief explanation of drawings]

FIG. 1 is an explanatory diagram illustrating problems that become noticeable when considering a configuration in which processing within the device is divided and processing results are sequentially transferred, and FIG. FIG. 3 shows an explanatory diagram illustrating one embodiment of the process according to the present invention, which corresponds to or corresponds to the illustrated explanatory diagram in FIG. 1. In the figure, 1 is the analysis section, 2 is the recognition section, 3 is the input/output management section, and 4
1.6 and 8.10 represent processors, respectively. Patent applicant Hiroshi Mori (1 other person), agent patent attorney of Fujitsu Ltd.

Claims (1)

[Claims] Performing frequency analysis of the input voice to generate a plurality of time series feature information, and determining the category of the input voice by matching with the standard feature time series stored in a dictionary, In a speech recognition device configured to output the recognition result to an input/output device, an analysis unit that generates the plurality of time series feature information, a recognition unit that determines the category, and an input/output manager that outputs the above. The analysis section performs /f-ebrining processing to transfer the results obtained by the analysis section to the recognition section and the results obtained by the certification section to the human/output management section. comprises a monitoring means for monitoring the length of the silent period of the input voice, and at the start of the silent period, transfers the time-series characteristic information based on the processing results up to that point to the recognition unit, The recognition unit is configured to notify the recognition unit whether or not the silent period is an inter-word silent period, and if it is not an inter-word silent period, transfer the time-series feature information again in the next silent period, and the recognition unit The unit is characterized in that it is configured to start the recognition process based on the received time-series feature information, and determine whether to continue or stop the drawn recognition process at the time of notification corresponding to the inter-word silent period. Speech recognition device.