JPH096384A - Voice recognition device - Google Patents

Abstract

PURPOSE: To reduce the instability in the change with time of likelihood of a candidate word and to perform a more stronger word detection in a word spotting device detecting a specified word in vocalizing. CONSTITUTION: A voice detection part 101 performs frequency analysis of an input voice, and extracts a feature vector at every frame. A likelihood calculation part 102 performs the matching of the time sequence of the feature vector with a phoneme dictionary 104 and a word dictionary 105 to calculate the likelihood of respective candidate words at every frame. An automaton part 106 corrects the likelihood at every frame. In such a case, the present frame likelihood is corrected according to the likelihood with at least one frame between the present likelihood and the past likelihood. Thus, the instability in the secular change of the likelihood is eliminated.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus, and more particularly to a word for calculating the likelihood of a specific word for each constant time (frame) during utterance and detecting the specific word based on the calculated likelihood. The present invention relates to a spotting device.

[0002]

2. Description of the Related Art Conventionally, as a technique for detecting a specific word in utterance, the IEICE Transactions, '84 / 10, Vo.
l. J67-D No. 10 p. 1242-1249
"Continuous Speech Recognition Algorithm by Extended Continuous DP Method"
(Seiichi Nakagawa), which is known as a word spotting device.

FIG. 8 is a block diagram showing an outline of this word spotting device, and shows a speech analysis unit 10 from the time of utterance.
1, the voice analysis is performed, the likelihood of the specific word for each frame is used as it is as a detection reference value, and the result output unit 103 performs detection by comparison with the detection threshold value.
The likelihood calculator 102 uses the phoneme dictionary 104 and the word dictionary 10
5. Likelihood calculation is performed for each frame with reference to FIG.

In this method, the likelihood calculated independently for each frame is used as it is as a detection reference value, and detection is performed by comparison with a predetermined detection threshold value.

[0005]

In such a conventional method, since the likelihood of each frame is used as it is as a detection value, the likelihood is unstable for a word that changes with time. For example, if the likelihood repeats up and down around the detection threshold, the same word may be detected multiple times in succession, and there is also the problem that a momentary increase in likelihood causes erroneous detection. .

An object of the present invention is to reduce the instability of the likelihood of a candidate word over time by reducing the adverse effect of using the likelihood of each frame as a detection reference value as it is, and to achieve more robust word detection. It is to provide a voice recognition device capable of performing.

[0007]

According to the present invention, the speech recognition for calculating the likelihood of a specific word for each constant time (frame) from the utterance and detecting the specific word based on the calculated likelihood. A speech recognition apparatus, comprising: a likelihood correction means for correcting and deriving the likelihood of each frame based on the likelihood of the current frame and the likelihood of at least one frame in the past. Is obtained.

[0008]

With the present invention, the likelihood of each frame is not used independently, but the likelihood of temporal change of the likelihood is reduced by referring to the likelihood of the current frame and the likelihood of at least one frame in the past. Is.

[0009]

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

FIG. 1 is a schematic block diagram of an embodiment of the present invention, and the same portions as those in FIG. 8 are designated by the same reference numerals. FIG.
In, the voice analysis unit 101 analyzes the frequency of the input voice, and extracts the feature vector for each fixed time (frame synchronization). Likelihood calculator 102 is speech analyzer 1.
The likelihood of each candidate word in each frame period is calculated by matching the time series of the feature vector from 01 with the phonological dictionary 104 and the word dictionary 105, and outputs the likelihood to the automaton unit 106.

The automaton section 106 is a likelihood calculation section 102.
The likelihood output by the above is received, the likelihood of each candidate word for each frame cycle is corrected, and the corrected likelihood is output to the output unit 103. The output unit 103 receives the likelihood of each candidate word, compares it with a predetermined threshold value, and outputs it as a detected word of a word having a likelihood of exceeding the threshold value for each frame period.

The difference from the conventional FIG. 8 is that the automaton unit 1
This is a point of correcting the likelihood calculated by the likelihood calculating unit 102 according to 06, and an example of likelihood correction in the automaton unit 106 will be described below. The automaton unit 106 shown here receives the likelihood for each frame output from the likelihood calculation unit 102 as an input likelihood, performs state transition by each case determination each time the likelihood is received, and subsequently, when each case is established. Processing shall be performed.

FIG. 2 is a diagram showing the first embodiment. The automaton in this example is an automaton whose initial state is s1, and the likelihood input in the state s1 is output as it is. At this time, if the likelihood is B (detection threshold) or more, the state transits to the state s2, and if not, the state remains at the state s1. The state s2 is a state in which the likelihood of the immediately preceding frame is B or more, and when the input likelihood is B or more, the state remains in the state s2 and the modified likelihood of less than B is continuously output. If not, the input likelihood is directly output as the modified likelihood and the state returns to the state s1.

In this automaton, a certain value less than the detection threshold is output as the modified likelihood only when both the likelihood in the current frame and the likelihood in the frame immediately before the current frame exceed the detection threshold. Otherwise, the likelihood of the current frame is output as it is. The modified likelihood output using this automaton is equal to or higher than the detection threshold only in the frame in which the input likelihood is increased from less than the detection threshold to above the detection threshold.
Also, if the likelihood of the first frame is greater than or equal to the detection threshold,
This likelihood is output.

By using such modified likelihood, it is possible to prevent the phenomenon that the same word is continuously detected for several consecutive frames.

It should be noted that the value less than B as the modified likelihood in the transition 2 of FIG. 2 may be -∞ in extreme if it is a value smaller than B, but since it is not practical, for example, "B" −
A value of 1,0 "may be output as the modified likelihood.

In the case of using the automaton shown in FIG. 2, if the detection threshold value is raised from B to B '(B'> B) as shown in FIG. However, in the case of B ', the number of times of detection may increase, contrary to t2 to t4. Therefore, the second shown in FIG.
Examples of are conceivable.

In the automaton shown in FIG. 4, the operation in the state s2 of the automaton shown in FIG. 2 is different.
In the state s2 of this automaton, the input likelihood is B2 (B2
In the case of <B) or more, the value remains at s2 and a value less than B is output as the modified likelihood. If not, the input likelihood is directly output as the modified likelihood, and the state is returned to the state s1. B2 is a detection activation threshold.

In this automaton, in addition to the detection threshold value, a detection activation threshold value that is less than the detection threshold value is determined. Even if the detection activation threshold value is once exceeded, the likelihood of exceeding the detection threshold value is not output until the detection activation threshold value is exceeded. . As a result, multiple detections are not performed while the likelihood remains above the detection validation threshold. Therefore, even if the detection threshold is increased as shown in FIG. 5, by setting the detection validation threshold to the same value as the original detection threshold, it is possible to prevent the phenomenon that the number of detections increases.

In other words, when the calculated likelihood of the current frame is equal to or higher than the detection threshold B and the likelihood of equal to or higher than B is not output in the past frame, the detection validation threshold B2 smaller than B is set. Unless there is a frame that outputs a likelihood that is lower than the previous frame and a frame that outputs a likelihood of B or more up to the previous frame thereafter, B
The above value is the modified likelihood (for example, the likelihood Lf of the current frame).
In other cases, a value less than B is derived as the modified likelihood.

FIG. 6 shows a third embodiment. The automaton shown in FIG. 6 is an automaton whose initial state is s1. When the input likelihood is less than B, it stays in the state s1 and is below the detection threshold. A value less than the detection threshold is output with (Lf in this example) as the modified likelihood.

In the state s2, if the input likelihood is B or more, it remains as it is. At this time, the state n is past n
A value equal to or greater than the detection threshold (Lf in this example) is output as the modified likelihood only when the number of frames remains equal to or more than the frame, and a value less than the detection threshold is output as the modified likelihood otherwise. When the input likelihood is less than B in the state s2, the state returns to the state s1 and the value less than the detection threshold is output as the modified likelihood.

This automaton outputs a value equal to or higher than the detection threshold value only when all the likelihoods in the past n frames are equal to or higher than the detection threshold value. With this automaton, it is possible to reduce false detection due to a momentary increase in likelihood that should not be detected.

FIG. 7 is a diagram showing a fourth embodiment. The automaton shown in FIG. 7 is an automaton whose initial state is s1 and stays at s1 until the nth frame is reached and includes the current frame. The average value of the likelihoods of all past frames is continuously output. When the nth frame is reached, the average of the likelihoods of the past n frames including the current frame is output, and the process proceeds to s2. After that, the average value of the likelihood of the past n frames including the current frame is output for each frame.

By doing so, the output modified likelihood becomes a weighted average value of the likelihoods in the past n frames including the current frame, so that the likelihood of the candidate word is smoothed, and the moment of the word that should not be originally detected is It is possible to reduce the probability of erroneous detection due to the increase in the likelihood of.

[0026]

As described above, according to the present invention, the calculated likelihood of the current frame is corrected and output according to the likelihood of the current frame and at least one frame in the past. Reduces the instability of changes in the likelihood of words over time,
There is an effect that a word spotting device capable of performing more robust word detection can be configured.

[Brief description of drawings]

FIG. 1 is a schematic system diagram of word spotting according to an embodiment of the present invention.

FIG. 2 is a diagram illustrating an example of the automaton unit 106 in FIG.

FIG. 3 is a diagram for explaining a problem of the embodiment of FIG.

FIG. 4 is a diagram illustrating another embodiment of the automaton unit 106 in FIG.

FIG. 5 is a diagram for explaining the effect of the embodiment of FIG.

FIG. 6 is a diagram illustrating still another embodiment of the automaton unit 106 in FIG.

FIG. 7 is a diagram for explaining another embodiment of the automaton unit 106 in FIG.

FIG. 8 is a schematic system diagram of a conventional word spotting device.

[Explanation of symbols]

 101 voice analysis unit 102 likelihood calculation unit 103 result output unit 104 phoneme dictionary 105 word dictionary 106 automaton unit

Claims (6)

[Claims] 1. A speech recognition apparatus for calculating the likelihood of a specific word for each constant time (frame) during utterance, and detecting the specific word based on the calculated likelihood. A speech recognition apparatus comprising a likelihood correction means for correcting and deriving a degree based on a likelihood in a current frame and a likelihood in at least one frame in the past. 2. The likelihood correction means derives a likelihood less than the threshold as a corrected likelihood when the likelihoods of the current frame and the frame immediately before the current frame are both equal to or larger than a predetermined threshold, and in other cases. 2. The speech recognition apparatus according to claim 1, wherein the likelihood of the current Reem is derived as a modified likelihood. 3. The speech recognition apparatus according to claim 2, wherein when the likelihood of the first frame is equal to or more than the threshold value, the likelihood is derived as a modified likelihood. 4. The likelihood correction means is a case where the calculated likelihood of the current frame is equal to or higher than a first threshold and the likelihood of equal to or higher than the first threshold is not output in a past frame, A frame that outputs a likelihood that is smaller than a second threshold and that is smaller than a first threshold exists before the immediately preceding frame and thereafter outputs a likelihood that is equal to or greater than the first threshold up to the immediately preceding frame. The value greater than or equal to the first threshold value is used as the modified likelihood unless it exists, and the value less than the first threshold value is derived as the modified likelihood in other cases. Speech recognizer. 5. The likelihood correction means derives a value equal to or more than the threshold value as a correction likelihood only when all likelihoods in a fixed number of past frames including the current frame are equal to or more than a predetermined threshold value, and otherwise. In the case of, the speech recognition device according to claim 1, wherein a value less than the threshold value is derived with a modified likelihood. 6. The likelihood correction means is configured to derive a weighted average value of likelihoods in a fixed number of past frames including the current frame as the corrected likelihood of the current frame. 1. The voice recognition device according to 1.