JPH02105200A - Speech recognizing device - Google Patents

Abstract

PURPOSE:To recognize a continuous voice more securely by totalizing a constant number of similarity values which continue from a current point to a past point at every identification label and extracting the identification label having the highest similarity among the totalization results. CONSTITUTION:A 1st arithmetic means 200 makes the continuous voice which is inputted from an input means 100 into feature parameters at every constant interval of time and calculates distances from plural standard parameters whose voice identification labels are already made clear to determine the similarity values of the identification labels. A 2nd arithmetic means 300 totalizes a constant number of continuous similarity values which are determined by the 1st arithmetic means 200 from the current point to the past point at every identification label. Then an extracting means 400 extracts the identification label corresponding to the totalization result having the highest similarity and output is as a speech recognition result. Consequently, the variation part of the voice can be recognized as the same label as a label which is recognized so far and the recognition probability of the continuous voice can be increased.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a speech recognition device that converts input speech into an identification label indicating the utterance of the speech.

(Conventional technology) Generally, a speech recognition device stores feature parameters extracted from speech in advance with an identification label indicating the utterance of the speech, and selects an identification label whose feature parameters are most similar to the input speech. Labels are extracted and output as speech recognition results. Among conventional speech recognition devices of this type, there is a known device that performs speech recognition at regular intervals and uses the speech recognition results to recognize phonemes. However, in the audio,
When there are 8 lines of utterance from one phoneme to the next, there is a phoneme in which the characteristics of the voice change at the boundary part (referred to as a transition area), and conventional speech recognition devices of this type are difficult to recognize in this transition area. There was a problem in that incorrect knowledge of identification labels had a negative effect on phonological recognition.

This point will be explained in detail.

(Problem to be solved by the invention) FIG. 7 shows a general speech recognition processing procedure of a conventional device.

In Fig. 7, for example, the voice uttered into the microphone, ``°°Ao°'', is determined by frequency analysis of the audio signal from the microphone. ). The distance between this feature parameter and each standard feature parameter stored in advance as described above is calculated, and by comparing the degree of similarity between this feature parameter and each standard feature parameter, the most similar one at that point is determined. The identification symbol (referred to as a label) is extracted.

The above-mentioned sound “ao′°” can be expressed symbolically as “
A label string of "Hehehe U to ULIA11UOOOOOO..." is obtained. However, in the phoneme °A °° and °゛o °° transition parts, the feature parameters change, so "
A” may be mistakenly recognized as “IJ”.

In consideration of such misrecognition, after a correction process is performed to convert the label sandwiched between two identical labels into that label, the number of consecutive label strings identified within a certain period of time is the highest. Extract the label as a phonological result. As a result, in the example shown in FIG. 7, the utterance "Ao" is incorrectly recognized as "UO" due to the influence of the erroneous recognition result of the transition part.

Conventional speech recognition devices that recognize continuous speech as described above have a problem in that they often misrecognize phonemes due to changes in the characteristic pattern of the transition portion between one phoneme and the next phoneme.

Therefore, the purpose of the present invention is to solve such problems,
An object of the present invention is to provide a speech recognition device capable of recognizing continuous speech more reliably.

(Means for Solving the Problem) In order to achieve such an object, the present invention includes an input means for inputting continuous speech, and a feature parameter for the continuous speech inputted from the input means at fixed time intervals. a first calculating means for calculating the distance between the parameter and each of the plurality of standard feature parameters for which the identification label of the speech is known in advance, and determining the degree of similarity of the plurality of identification labels for the continuous speech based on the calculation result; and a second calculation means for summing up a certain number of consecutive similarities from the current time determined by the first calculation means to the past for each identification label; The present invention is characterized by comprising an extraction means for extracting an identification label corresponding to the summation result having the highest degree of similarity based on the summation result, and outputting the extracted identification label as the current speech recognition result.

[For production]

In the present invention, a predetermined number of past similarities starting from the similarities obtained at each point in time are summed for each identification label by a second calculation means, and an identification label having the highest similarity among the total results is extracted. Extract sequentially by As a result, the similarity obtained at each point in time is smoothed over time, so for example, noise may be partially mixed into the input speech, the utterance of a part of the input speech may change, or the phonology of the input speech may be smoothed over time. Even when the voice changes, the changed part of the voice is recognized as the same label as the label recognized so far, and the probability of voice recognition increases.

〔Example〕

Embodiments of the present invention will be described in detail below with reference to the drawings.

The applicant of this application focused on the property of continuous speech in which the characteristic pattern of the transition part of continuous speech changes little by little from the characteristic pattern of the speech sampled up to the previous point in time. The calculated similarity of each standard feature parameter is aggregated over a certain period of time (referred to as a window) in chronological order, and based on the aggregated results, the identification label with the highest degree of similarity is determined as the recognition result at each point in time. It is something.

FIG. 1 shows the basic configuration of an embodiment of the present invention.

In FIG. 1, 100 is an input means for inputting continuous speech.

200 converts the continuous speech inputted from the input means into feature parameters at regular intervals, calculates the distance between the feature parameters and each of a plurality of standard feature parameters for which the identification label of the speech is known in advance, and It is a first calculation means that determines the degree of similarity of the plurality of identification labels with respect to the continuous speech based on the calculation results.

Reference numeral 300 denotes a second calculation means that totals a certain number of consecutive similarities from the present time to the past determined by the first calculation means for each of the identification labels.

400 extracts an identification label corresponding to the total result with the highest degree of similarity based on the total result of the similarity for each identification label summed by the second calculation means, and uses the extracted identification label as This is an extraction means that outputs the voice recognition result at the current time.

FIG. 2 shows a specific configuration of an embodiment of the present invention.

In FIG. 2, 10 is a microphone serving as an input means for inputting voice. An amplifier 11 amplifies the output of the microphone. 12 is an A/D converter, which converts the amplified output of the amplifier 11 into A/D. 13 is the first
It is a Fourier transformer that constitutes a part of the calculation means, and the A/
The output of the D converter 12 is Fourier-transformed, and the sound strength (power spectrum) for each frequency band is output as a sound characteristic parameter. The Fourier transformer 13 is an LSI (
Large-scale integrated circuits) are known.
A Fourier transform may be performed by a computer. It is also possible to use a bandpass filter instead of a Fourier transformer.

14 is the first. It is a computer system corresponding to the second calculation means and the extraction means, and the computer system 14
is connected to a keyboard 15, an external storage device (FDD) 16 using a floppy disk or the like, a printer 17, and a cathode ray tube display (CRT) 18. 15 is a keyboard that can input information; from the keyboard 15,
Input information for voice recognition mode instructions and standard feature parameter creation, modify recognition results, etc. The FDD 16 stores the created standard feature parameters and the code numbers of identification labels corresponding to the standard feature parameters as a table (referred to as a phoneme map). Also,
The FDD 16 includes a probability table 16 related to the present invention, which will be described later.
-1 is provided.

FIG. 3 shows an example of the configuration of the computer system 14 according to the embodiment of the present invention.

In this embodiment, in order to perform high-speed arithmetic processing, arithmetic processing is performed by a plurality of central processing units. In FIG. 3, 14-1 is a microprocessor (MPU), the MPU 14-1 labels input speech (speech recognition), and 14-2 is a memory.
2 stores control procedures executed by the MPU 14-1. 14-3 is a high-speed digital signal processor (o
14-4 is a memory;
Stores control procedures executed by the DSP 14-3. 14-
A personal computer 5 plays the role of continuing the operation of the entire system, including setting the recognition mode and displaying the recognition results.

Note that a large-sized computer may be used as the computer system 14, and the configuration may be configured according to the size of the device and the degree of arithmetic processing.

FIG. 4 shows the memory configuration of the probability table 16-1 shown in FIG. 2.

In FIG. 4, a probability table 16-1 stores probability information calculated at fixed time intervals from the voice input start time TO for each identification label.

This probability information is the result of calculating the distance between the input voice feature parameter and the standard feature parameter, expressed as a percentage for each identification label. Note that the distance calculation result itself may be used as this probability information.

The higher the value of this probability information, the more similar the feature parameters of the input voice and the feature parameters of its label are.

FIG. 5 shows a control procedure executed by the MPU 14-1 shown in FIG. The operation of the embodiment of the present invention will be explained with reference to the flowchart of FIG.

In FIG. 5, the sound input from the microphone 10 is converted into characteristic parameters for each time point by the Fourier transformer 13 at constant time intervals from input start time TO to TI, T2, . . . .

When the MPU 14-1 inputs this input feature parameter, it instructs the DSP 14-3 to calculate distances from each standard parameter (steps Sl to 3). When the MPU 14-1 receives the identification label and the distance calculation result for this label from the DSP 14-3, it converts it into probability information and writes it into the probability table 16-1 (see FIG. 4) (step S4). If we consider the same example as the utterance "Ao" explained in the description of the conventional example, probability information can be obtained such as, for example, 0.7" for the identification label "A", °゛02" for "i", etc. .

In this embodiment, the MPU 114-1 sums up the values of probability information for each identification label from the time T when the characteristic parameters of the input voice are input to the past four time points, and selects the identification label with the highest total result at the time T. Defined as the phoneme identification result. That is, the calculation formula is as follows.

Here, Pj(t): total probability of identification label j at time t,
(: t, j: probability of phoneme J at time t, Bk: weighting coefficient for the total number of probabilities, at: normalization coefficient at time t, ! = number of probabilities to be summed - 1 (length of time span) )
, In this example, an example in which I-4, Bk-1, and 8t-1 are set is explained.

Therefore, as shown in FIG. 6, this total result is also "0.7" in the order of identification labels "A", "I", and "0".
0.2 ” “0.0 ” and FDD16, MPl
+14-1 is written (step S5).

When the sum of probability information values at time TO is completed for all identification labels in this manner, the MPU 14-1 extracts the identification label with the highest total value and stores it in the FD016 (steps 56 to 57). ).

In this example, the identification label "A'" having a probability of 0.7 is stored as the identification result.

MPI+14-1 repeats these steps,
Probability information and probability information total values are stored in FD016, and at time T5, five probability values from times T1 to T5 are summed. Hereinafter, as shown in FIG. 6, the range to be totaled (referred to as a window) is sequentially moved.

Furthermore, the label with the highest number of identification labels among the identification labels obtained in this way is output as the phoneme recognition result. A conventionally known control procedure can be used for this purpose (step S8).

In continuous speech, when the utterance (sound M) changes, the feature pattern for each phoneme also changes little by little, so conventionally, the phonological label of the true standard feature parameter corresponding to the utterance was used as the second or third identification candidate. often appear.

Since the present invention also uses the distance calculation results within a predetermined period of time in the past from the time when the voice feature parameters are input, for voice recognition processing, even if at a certain time point T, the recognition result becomes the second candidate in the conventional method. However, if the identification results have been the first candidate multiple times in the past, the candidate is determined to be the first candidate at that time T.

For example, in the conventional device that selects only the first candidate for the utterance “ao” explained in the conventional example, “＾AAtl
The identification result of partially different phoneme labels such as "AtllJAUII00..." is obtained, but in the present invention, the same label is identified for every plural number of phoneme labels as shown in the label string "AAAAAAA 88 AUU to...°°". In other words, it is obtained as a smoothed label sequence.

As a result, as explained in the conventional example for phoneme recognition processing,
It is clear that there is no need to perform identification label correction processing, and the speech recognition probability increases.

The following can be considered as an application form of the present invention.

1) In this example, the recognition probability converted from the distance calculation result is used as the degree of similarity between the standard feature parameter and the characteristic parameter of the input speech. good. Therefore, in this case, the ranking values at each time point are summed for each identification label, and the identification label with the smallest value becomes the identification result.

2) In this embodiment, an example is shown in which all the similarities used for speech recognition are stored in the FDD 16, but registers and memories are prepared for the number of similarities to be added, and each time a new similarity is obtained, If the memory contents are updated frequently, it becomes possible to reduce the memory capacity of the speech recognition device.

3) In this example, identification labels are extracted using the past multiple similarity total results as an index, but in addition to the total results,
An average value of multiple degrees of similarity or a statistically smoothed value can be used as an index.

(Effects of the Invention) As explained above, in the present invention, a predetermined number of past similarities starting from the similarities obtained at each point in time are summed by the second calculation means for each identification label, and the total result is The identification label with the highest degree of similarity is sequentially extracted by the extraction means.As a result, the degree of similarity obtained at each point in time is smoothed over time. Even if the occurrence of a part of the input speech changes, or the phonology of the input speech changes, the negative effects caused by the changed part of the speech are removed and the correct label is recognized as a whole, thereby improving speech recognition. The probability increases.

[Brief explanation of drawings]

Figure 1 is a block diagram showing the basic configuration of an embodiment of the present invention, Figure 2 is a block diagram showing a specific configuration of an embodiment of the invention, and Figure 3 is a circuit diagram of the computer system shown in Figure 2. 4 is a memory map showing the memory configuration of the probability table 16-1 shown in FIG. 2, FIG. 5 is a flow chart showing the control procedure executed by the MPUI 4-1 shown in FIG. 3, and FIG. FIG. 7 is an explanatory diagram showing the process of calculating the similarity of the invention fruit color example. FIG. 7 is an explanatory diagram showing the speech recognition process of the conventional example. ! 0...Microphone, 14...Computer system, 14-1
... Microprocessor (MPIJ), 16
... Floppy disk device (FDD), 16-1
... Probability table. The trouble of 4 is the stone of 11 Nippon Marozo former Wara Agency envoy 1, which is the combination of the non-guardian and fugata series, -ku system)! Bisible;
6-; Please show T〆t Remap Duke 1.
T1 TO~T2 To-T3 Ding O~Eng 4A
O, 7152, 32, 43, 2+ 0.2 0.
3 0.4 0.4 0.4U O,0
0, 00, 00, 60, 8 uninvented piece 1) Calculate the color degree 7 Figure 6 D1~T5 T2 da 6--1 2.6 2.1 0.2 0. + 1.42・O ;

Claims (1)

[Claims] 1) An input means for inputting continuous speech, and a method for converting the continuous speech inputted from the input means into feature parameters at fixed time intervals, the feature parameters and the identification label of the speech being known in advance. a first calculation means for calculating the distance between each of the plurality of standard feature parameters, and determining the degree of similarity of the plurality of identification labels for the continuous speech based on the calculation result; a second calculation means for summing up a certain number of consecutive similarities from the present time to the past for each of the identification labels; 1. A speech recognition device, comprising: extracting means for extracting an identification label corresponding to the total result with the highest degree of similarity, and outputting the extracted identification label as the speech recognition result at the current point in time.