JPH0484197A - Continuous voice recognizer - 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] Industrial Application Field The present invention relates to a continuous speech recognition device that obtains recognition results using evaluation values expressed by linear combinations of acoustic reliability and connection reliability that are each variably weighted. With the development of speech recognition technology, continuous speech recognition devices are about to be put into practical use in various fields. One of the practical problems with this is that if there is a part that is uttered unclearly in the input continuous speech, this part will be misrecognized (＼ An incorrect recognized sentence will be output. In order to solve the above problems, conventional continuous speech recognition devices do not simply connect segments recognized with the highest acoustic reliability to obtain sentence-by-sentence recognition results. The evaluation function is set as a linear combination of the acoustic reliability and connection reliability of the segment, and the one with the highest evaluation value is taken as the recognition result, so even if the acoustic reliability of the segment is low, the previous and subsequent segments If the grammatical connection (connection reliability) is high, the evaluation value will be high, and even if there are unclearly uttered parts in the input continuous speech, it will be possible to recognize them correctly.

Hereinafter, referring to FIG. 5, explanation will be given by taking as an example the conventional continuous speech recognition device as described above, in which the fragments are words.

Figure 3 Table C Block diagram of conventional continuous speech recognition device
FIG. 4 is a flowchart of word lattice generation, and FIG. 5 is a flowchart showing processing at the connection section. In Fig. 3, 1 is a signal input terminal, 2 is an analysis section, 4 is a feature parameter storage hole, 5 is a collation section, 6 is a word lattice storage suit, 8 is a prediction section, 9 is a connection section, 10 is a recognition result output terminal, and 12 is a switch. The operation of the speech recognition device configured as described above will be explained below.

First, standard voice registration will be explained with reference to FIG. The standard voice obtained is input to the analysis section 2, feature parameters are calculated for each frame, and the feature parameters are registered in the feature parameter storage section 4. Then, the standard speech is input and the above process is repeated to complete the registration of all the words to be recognized.

Next, the time of recognition will be explained with reference to FIGS. 3 and 4. First, operate the switch 12 so that the output of the analysis section is input to the verification section, process 21) so that the signal to be recognized is input from the signal input terminal 1 in the same way as when registering, and the analysis section 2.
Calculate the feature parameters for each frame (process 22
). Next, a comparison section 5 performs a comparison between the standard voice and the input signal.

First, initialization is performed by setting the frame number to 1 and the word number to 1 (processes 23 and 24). Then, the frame with the frame number is used as the starting point to match the word to be recognized with the word number (processing 27), and if the degree of similarity is greater than or equal to the determination threshold (processing 28), the word at this time is used as a recognition segment. The candidate similarity is used as the acoustic reliability and is output to the word lattice storage unit along with the recognition start and end points (process 29). After that, add 1 to the word number (process 30) and move on to processing for the next word to be recognized. When the above processing is completed for all the words to be recognized (process 26), add 1 to the frame number.
is added (process 31), and the same process is performed using the next frame as the starting point. When the above processing is completed for all frames and all words to be recognized (processing 25)
, the processing in the collation unit 5 is completed. This processing identifies the names of recognized word candidates that may exist in the input continuous speech as a word lattice and their starting positions. The end position acoustic reliability is recorded and output to the word lattice storage section 6.The prediction section 8 also uses grammar or statistical information to find words that can follow the recognition word candidate input from the connection section 9.Next word Next, the processing of the connection unit 9 which operates to output the connection reliability (expressed by probability or the like) as a candidate to the connection unit 9 will be explained with reference to FIGS. 3 and 5.

The connection section connects the recognition word candidate names input from the word lattice storage section 6 to generate multiple recognition result candidates.
The one with the highest evaluation value is the recognition result and L
The recognition result is output to the output terminal 8. To obtain recognition result candidates, first process the recognition result as an empty string 1)
, initialization is performed (process 2) by setting the recognition intermediate result end position, which is the end position of the last recognized word candidate in the recognition intermediate result, to 0 (frame number). Next, if there is a recognized word candidate (possible to follow) that satisfies the start position condition of recognition intermediate result end position - gap ≦ start position ≦ recognition intermediate result end position + gap (formula 1, 1) (process 4), The recognized word candidate name is connected to the end of the character string in the recognition intermediate result (processing 5). Here, gap is a constant for dealing with the detection error between the start end position and the end position in the processing in the matching section. Thereafter, the recognition intermediate result end position is updated using the following equation 2 (processing 6).

Recognition intermediate result end position = Recognition intermediate result end position + (recognition word candidate end position -
Starting position of recognition word candidate) (Equation 2) Then, assuming that the recognition word candidate connected in process 34 is the first recognition word candidate from the beginning of the sentence, the connection unit 4 sends the i-1st recognition word candidate from the beginning of the sentence to the prediction unit. , the next word candidate (word candidate that could be the i-th word from the beginning of the sentence) and its connection reliability are obtained from the prediction unit (processing 7). Then, using these, the evaluation value is updated according to the following equation 3 (processing 9). At this time, the word string (W
l, W2. The evaluation function h(Wi) of the recognition intermediate result, which is composed of , , , Wi), is expressed as follows. be.

Here g(Wi) is the acoustic reliability of word Wi, f(W
i-1, Wi) is the connection reliability from word Wi-1 to word w1, and α and β are weights (constants). Thereafter, processes 4 to 9 are repeated until the condition in process 4 is no longer satisfied. After that, input audio frame length - gap ≦ recognition intermediate result end position ≦
Input audio frame length + gap (Formula 1, 2) Process to determine the end position of the mid-recognition result under the condition 10)
, if this condition is met, the intermediate recognition result at this time is saved as a recognition result candidate along with its evaluation value (
Processing 13). The above procedure for obtaining intermediate recognition results is performed for the entire word lattice that is input to the connection unit 9. The recognition result is output from the output terminal 11 as a short answer to the problem that the invention attempts to solve.In the above configuration, the weights (α and β in equation 3) applied to acoustic reliability and connection reliability are fixed. In addition, the above configuration had the problem that the two-way reliability was not efficiently reflected in the evaluation value, and the introduction of the reliability had no effect on improving the recognition rate. If there is a silent section, the recognition result may not be output because the recognition segment candidates cannot be connected, or if the presence of a silent section is assumed in advance, Equation 1.1
If the gap value is increased, a huge number of recognition result candidates will be output. The weights (α and β in Equation 3) applied to the acoustic reliability and connection reliability are changed according to the amount of information for segment prediction, the silent time in the input speech, or both, and each reliability is faithfully adjusted to the recognition rate. It is an object of the present invention to provide a continuous speech recognition device that leads to an improvement in the number of frames of an input signal. an analysis section that detects feature parameters, a verification section that compares the pressure of the analysis section with the feature parameters of each segment of the standard signal and outputs recognition segment candidates and their acoustic reliability; Recognition is performed by connecting the prediction unit that outputs the next segment candidate predicted from the recognition intermediate result, its connection reliability, and the amount of information of the predicted next segment candidate, and the recognition unit candidate that is the output of the matching unit. It is composed of a connection section that outputs the results and also outputs the results during recognition to the prediction section.

Continuous speech recognition device according to the second invention: an analysis section that detects feature parameters for each frame of an input signal; a detection section that detects silent sections of the input signal; and an output of the analysis section and each segment of a standard signal. a matching unit that outputs a recognition unit candidate and its acoustic reliability by comparing it with the feature parameters of the unit, and outputs a next unit candidate and its connection reliability that are predicted from the recognition intermediate results input from the connection unit. It is composed of a prediction section and a connection section that connects recognition unit candidates output from the matching section to output recognition results and outputs intermediate recognition results to the prediction section.

The continuous speech recognition device according to the third invention includes an analysis section that detects feature parameters for each frame of an input signal, a detection section that detects a silent section of the input signal, and an analysis section that detects feature parameters for each frame of an input signal, a detection section that detects a silent section of the input signal, and an output of the analysis section and a standard signal for each segment. A matching unit that outputs recognition segment candidates and their acoustic reliability by comparing them with feature parameters, and a next segment candidate, its connection reliability, and its next segment predicted from the recognition intermediate results input from the connection unit. It is composed of a prediction unit that outputs the information amount of the candidate, and a connection unit that connects the recognition segment candidates output from the matching unit to output a recognition result and outputs an intermediate recognition result to the prediction unit.

Continuous speech recognition device of the first invention (translation) The analysis section detects the feature parameters for each frame of the input signal, and the matching section compares the output of the analysis section with the feature parameters for each segment of the standard signal to generate recognition elements. The segment candidate and its acoustic reliability are output, and the prediction unit connects the next segment candidate predicted from the recognition intermediate result input from the connection unit, its connection reliability, and the predicted information amount of the next segment candidate. At the connection section, the acoustic reliability input from the matching section and the connection reliability input from the prediction section are each weighted according to the amount of information of the next segment candidate input from the prediction section. Recognition segment candidates are connected using the evaluation value expressed by to obtain recognition results, and intermediate recognition results are output to the prediction unit.

Continuous speech recognition device of the second invention (an analysis section detects feature parameters for each frame of the input signal, a detection section detects silent sections of the input signal, and a collation section compares the output of the analysis section with a segment of the standard signal) The prediction unit outputs the recognition unit candidate and its acoustic reliability by comparing it with the feature parameters for each unit, and the prediction unit outputs the next unit candidate and its connection reliability predicted from the recognition intermediate results input from the connection unit. At the connection section, the acoustic reliability input from the matching section and the connection reliability input from the prediction section are expressed as a linear combination of each weighted according to the time length of the silent section input from the detection section. The continuous speech recognition device of the third invention has the following characteristics for each frame of the input signal: On detecting the parameters, the detection section detects the silent section of the above input signal, and the matching section compares the output of the analysis section with the characteristic parameters of each segment of the standard signal to output recognition segment candidates and their acoustic reliability. The prediction unit outputs the next segment candidate predicted from the intermediate recognition results input from the connection unit, its connection reliability, and the amount of information of the next segment candidate to the connection unit, and outputs the information input from the matching unit at the connection unit. A linear method in which the acoustic reliability and the above-mentioned connection reliability input from the prediction unit are respectively weighted according to the amount of information of the next unit candidate input from the prediction unit and the time length of the silent section input from the detection unit. Recognition segment candidates are connected using evaluation values expressed by combinations to obtain recognition results, and intermediate recognition results are output to the prediction unit.

Embodiments Hereinafter, embodiments of the first, second and third inventions will be explained with reference to Table 1. Force exchange describing the device The evaluation value calculation process in this embodiment is effective even when the amount of information of the next segment candidate or the time length of the silent section is not used as input, so the invention according to claim 1 This will also be described as an embodiment of the invention set forth in claim 2.

FIG. 1 is a block diagram of a continuous speech recognition device using words as fragments in one embodiment of the present invention. ■ is the signal input terminal, 2 is the analysis section, 3 is the detection section, 4 is the feature parameter storage, 5 is the matching section, 6 is the word lattice storage, 7 is the weighting section (1), 8 is the prediction section, 9 is the connection section, 10 is the weight Attached is a part (2), 11 is a recognition result output terminal, and 12 is a switch. The operation of the speech recognition device configured as described above will be explained below. First, when registering a standard signal, switch 11 is operated so that the output of the analysis section is input to the feature parameter storage section.
The standard signal input from the signal input terminal 1 is input to the analysis section 2, which calculates feature parameters such as LPC cepstrum for each frame, and inputs it to the feature parameter storage section 4.The standard signal is input and the above processing is recognized. Repeat for all the words to be recognized End the registration 4 In addition, in the example (Table) A statistically determined connection probability (probability that each next word candidate appears after the previous word appears) is registered as this connection reliability.

In addition, in the embodiment, the verb plexity e (Wi) is used as the amount of information of the next word candidate, calculated by the following formula O, and registered in the prediction unit 8 (.

In this case, p (Wi-1, Wi) is the connection reliability of the word Wi following the word Wi-1. An example of this registration is shown in Table 1.

Table 1 This shows Wi-1 as the previous word Wi as the next word candidate, and when the recognition word candidate is input from the connection part 9, the prediction unit 8 selects the words that can follow the recognition word candidate and their connections. It operates to output the probability and perplexity to the connection unit 9 as the information amount of the next word candidate, persistence probability, and next word candidate, respectively.

Next, the recognition process will be explained with reference to Figures 1 and 4. First, operate the switch 12 so that the output of the analysis unit is input to the verification unit, Processing for inputting signals to be recognized 21), analysis section 2
Calculate the feature parameters for each frame (process 22
). Next, a comparison section 5 performs a comparison between the standard voice and the input signal.

First, initialization is performed by setting the frame number to -1 and the word number to 1 (processes 23 and 24). Then, the frame with the frame number is used as the starting point to match the word to be recognized with the word number (processing 27), and if the degree of similarity is greater than or equal to the determination threshold (processing 28), the word at this time is used as a recognition result candidate. The degree of similarity is used as the degree of acoustic reliability and is output to the word lattice storage unit along with the start and end points of recognition (processing 29). After that, add 1 to the word number (process 30) and move on to processing for the next word to be recognized.4 When the above processing is completed for all the words to be recognized (process 26), add 1 to the frame number.
is added (process 31), and the same process is performed using the next frame as the starting point. When the above processing is completed for all frames and all words to be recognized (processing 25)
, the processing in the matching unit 5 is ended. Through this processing, the names of recognized word candidates that may exist in the input continuous speech as a word lattice, their starting position, ending position, and acoustic reliability are recorded and output to the word lattice storage unit 6. However, the word lattice generation method described above is the same as the conventional example.4 The input signal is also input to the detection unit, and the power of the input signal is calculated for each frame. Determine that there is no sound 4 A series of silent frames is defined as a silent section, and its start position and end position are set as one set and output to the connection unit 9.

Next, the processing of the connecting portion 9 will be explained with reference to FIGS. 1 and 2.

The connection section connects the recognition word candidate names input from the word lattice storage section 6 to generate multiple recognition result candidates.
Among these boards, the one with the highest evaluation value is recognized as the recognition result.
l Output to recognition result output terminal 8 To obtain recognition result candidates (process 1 by treating the recognition result as an empty string), output the recognition result candidate that is the end position of the last recognized word candidate in the recognition result. Set the result end position to 0 (frame number)
(Process 2) Initialize t, X, and input position information of the silent section such as the start position and end position of the silent section from the detection unit (Process 3). Next, if there is a recognition word candidate (possible to follow) that satisfies the start position condition of (Equation 1, 1) (processing 4), The recognized word candidate name is connected to the end of the character string in the recognition intermediate result (processing 5). Here, gap is a constant for dealing with a detection error between the start position and the end position in the processing in the matching section. However, if there is no recognized word candidate that satisfies Equation 1 (Processing 4), and input audio frame length - gap ≦ recognition intermediate result end position ≦ input audio frame length + gap (Equations 1, 2) is not satisfied (Processing 10), Recognition intermediate result end position = gap ≦ start position of silent section ≦
If the recognition intermediate result end position + gap (Formula 1, 3) is satisfied (processing 11), it is determined that a silent section exists L. The recognition intermediate result end position is extended by the time length of the silent section (processing 12) and the process returns to process 4. Return. Formula 1. If condition 3 is not satisfied (process 11), it is determined that there is no subsequent word and that the recognition results up to that point are incorrect, and the process is terminated.

This method allows processing to be performed without changing the gap value even if there is a silent section in the input voice due to the speaker's breathing.

When there is a next word candidate that can be followed in process 4 (process 5 to connect the next word candidate name to the end of the character string in the table recognition intermediate result), update the recognition intermediate result end position using the following formula 2. Perform (process 6).

Recognition intermediate result end position = Recognition intermediate result end position + (end position of recognition word candidate - start position of recognition word candidate)
(Formula 2) and processing 5
Assuming that the recognized word candidate connected in is the first recognized word candidate from the beginning of the sentence, the connecting unit 4 sends the 1-1st recognized word candidate from the beginning of the sentence to the prediction unit, and selects the next word candidate (word that can be the first word from the beginning of the sentence). candidate) and its connection reliability from the prediction unit (
Processing 7).

The prediction unit also obtains the information amount of the next word candidate (C1th word candidate) from the i-1st recognized word candidate from the beginning of the sentence.
Then, using these, the evaluation value is updated according to the following equation 3 (processing 7). At this time, the word string (Wl, W2
．． The evaluation function h (Wi) of the recognition intermediate result consisting of , , , Wi) can be expressed as follows: + b−f (Wi−1, Wi) (Formula 3) where h (WO) = O.

a”l”　・r　・e　(Wi).

bcl: δ/ (τ ・ e (Wi)) where g (Wi) is the acoustic reliability of the word Wi that the word lattice storage unit 6 outputs to the weighted unit (1) 7, and the weighting is In addition, f (Wi-1, Wi) is weighted by the prediction unit 8 and outputted to the unit (2) 10 from the word Wi-1. W
It is the connection reliability to i, and the weighting is given by weight b in part (2) 10 and input to the connection part 9. The weight a of the acoustic reliability is the information of the next word candidate input from the prediction part 8. The connection reliability weight b is a function inversely proportional to the amount of information of the next word candidate input from the prediction unit 8. In addition, τ is the time length of the silent section (end position of the silent section - start position of the silent section) when there is a silent section immediately before the recognition word candidate Wi.
) has a minimum value set so that b does not become infinite. Note that γ and δ are constants (fixed). With this method, when the amount of information of the next word candidate is large, priority is given to acoustic reliability, and when it is small, priority is given to connection reliability. Therefore, even if a series of word strings that tend to be uttered indistinctly (such word strings have a small amount of information for next word candidates and silent intervals tend to be difficult to exist) are input and the acoustic reliability is low, It is possible to improve the recognition rate by prioritizing connection reliability and increasing the evaluation value. Thereafter, processes 4 to 9 are repeated until the condition of process 4 is no longer satisfied.

Thereafter, if Equations 1 and 2 are satisfied (process 10), the intermediate recognition result at this time is saved as a recognition result candidate together with its evaluation value (process 13). The above procedure for obtaining intermediate recognition results is performed for the entire word lattice input to the connection unit 9 (X), and the recognition result candidate with the highest evaluation value is recognized among them. The result is output from the recognition result output terminal 11.

As described above, the present embodiment relates to the invention described in claim 3. When the time length of the silent section is not used in the power evaluation value calculation process (as described in claim 1), or when the amount of information of the next segment candidate is It is effective even when not using (as described in claim 2).

Effects of the Invention Continuous speech recognition device of the first invention (in order to change the weight given to acoustic reliability and connection reliability according to the amount of information of the next segment candidate, both reliability levels are efficiently reflected in the evaluation value. Even if the acoustic reliability of the input signal is low, if the amount of information of the next segment candidate is small, it is possible to give priority to the connection reliability and increase the evaluation value. It becomes possible to obtain correct recognition results.

Continuous speech recognition device of the second invention (to change the weight to be applied to the acoustic reliability and the connection reliability according to the time length of a silent section in input speech, and to efficiently reflect both reliability in the evaluation value. In other words, even if the acoustic reliability of the input signal is low, if there are no silent sections, it is possible to give priority to the connection reliability and increase the evaluation value. Atsushi: Correct recognition results can be obtained.

Furthermore, if there are silent sections such as breath breaks in the input speech, it is possible to connect recognition unit candidates without significantly increasing the number of recognition result candidates, and no recognition results will be output.

Continuous speech recognition device ζ of the third invention: To change the weight to be applied to acoustic reliability and connection reliability according to the amount of information of the next segment candidate and the time length of the silent section in the input speech.1.
, not only can the effect of the second invention be obtained, but also the quality of the input signal is improved because it incorporates the characteristics of the voice, which is the input signal, such that the amount of information of the next segment candidate is proportional to the time length of the silent section. By doing so, the recognition rate can be improved.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a speech recognition device according to an embodiment of the present invention. FIG. 2 is a flowchart of processing at a connection section in an embodiment of the present invention. The figure is a flowchart of word lattice generation in the conventional example, and FIG. 5 is a flowchart of processing at the connection section in the conventional example. 1...Input terminal, 2...Analysis test 3...Detection test 4...Characteristic parameter storage and withdrawal 5...Verification edge 6
...Word lattice storage s, 7...Weighting is part (1)
, 8... Prediction test 9... Connection part 10... Weighting part (2), 11... Output terminal, 12... Switch. Name of agent: Patent attorney Shigetaka Awano Haka 1 base 1 Figure / Figure 12 Shoshushincho/f (Approved Nb Keifu Gorge Supplement 9) Figure / Front figure

Claims (3)

[Claims] (1) An analysis section that detects feature parameters for each unit time (hereinafter referred to as a frame) of the input signal, and compares the output of the above analysis section with the feature parameters for each fixed time period (hereinafter referred to as a segment) of the standard signal. A matching unit that outputs recognition segment candidates and their similarity (hereinafter referred to as acoustic reliability); and a unit that connects the recognition segment candidates that are the output of the matching unit to output recognition results, and a prediction unit that predicts intermediate recognition results. the next segment candidate predicted from the above recognition intermediate result inputted from the above connection part, the reliability that the segment will appear (hereinafter referred to as connection reliability), and the predicted next segment. and the prediction unit that outputs the information amount of the candidate to the connection unit, and the connection unit outputs the acoustic reliability input from the verification unit and the connection reliability input from the prediction unit. The recognition unit candidates are connected using evaluation values expressed by linear combinations that are each weighted according to the amount of information of the next unit candidate inputted from the prediction unit to obtain a continuous speech recognition result. Continuous speech recognition device. (2) An analysis section that detects feature parameters for each frame of the input signal, a detection section that detects silent sections of the input signal, and a comparison between the output of the analysis section and the feature parameters for each segment of the standard signal. a matching unit that outputs recognition segment candidates and their acoustic reliability; and a connection unit that connects the recognition segment candidates that are output from the matching unit to output recognition results and outputs intermediate recognition results to the prediction unit. and the prediction unit that outputs the next segment candidate and its connection reliability predicted from the recognition intermediate result inputted from the connection unit to the connection unit, and the connection unit Using an evaluation value expressed by a linear combination of the input acoustic reliability and the connection reliability input from the prediction unit, each weighted according to the time length of the silent section input from the detection unit. A continuous speech recognition device characterized in that a continuous speech recognition result is obtained by connecting the recognition unit candidates. (3) An analysis section that detects feature parameters for each frame of the input signal, a detection section that detects silent sections of the input signal, and a comparison between the output of the analysis section and the feature parameters for each segment of the standard signal. a matching unit that outputs recognition segment candidates and their acoustic reliability; and a connection unit that connects the recognition segment candidates that are output from the matching unit to output recognition results and outputs intermediate recognition results to the prediction unit. and the prediction unit that outputs the next segment candidate, its connection reliability, and the information amount of the next segment candidate predicted from the recognition intermediate result inputted from the connection unit to the connection unit. , in the connection section, the acoustic reliability input from the matching section and the connection reliability input from the prediction section are combined with the information amount of the next segment candidate input from the prediction section and the detection. Continuous speech, characterized in that continuous speech recognition results are obtained by connecting the recognition segment candidates using evaluation values expressed by linear combinations each weighted according to the time length of a silent section input from the section. recognition device.