JPH11288293A - Voice recognition device and storage medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To realize a voice recognition device and a storage medium which improve the recognition precision. SOLUTION: A recognition start time setting part 14 sets a prescribed point of time within a range after the start time of an input signal and before the start time of a sound section as the recognition start time. Thus, the problem is resolved that erroneous recognition is caused by non-coincidence between the start time of the sound section and that of recognition due to an influence of the noise like breathing or a rip noise in a method which takes the start time of the sound section as that of recognition. A discrimination part 16 refers to a parameter 22 of a sound hidden Markov model and a parameter 24 of a silence hidden Markov model to perform discrimination by the hidden Markov model method. Thus, the recognition precision is not affected by breathing or a rip noise.

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 for recognizing speech and a storage medium storing a computer program for the speech recognition apparatus to perform speech recognition, and can improve the accuracy of speech recognition. Things.

[0002]

2. Description of the Related Art Conventionally, as the above-mentioned speech recognition device, for example, one having a configuration shown in FIG. 4 is known. The voice recognition device 40 includes a sound section start point detection section 42, a sound section end point detection section 12, a storage section 46, and an identification section 44. The voiced segment start time detection unit 42 receives the digitized signal input by the voice input device 30 as an input signal, detects the power of the input signal, and detects the start time of the voiced segment (hereinafter referred to as the start time). , A sound segment start time). The sound section end point detection unit 12 detects the power of the input signal or the like and sets the end point of the sound section (hereinafter referred to as the end point of the sound section). Further, the storage unit 46 stores the reference parameter 22 of the voiced hidden Markov model corresponding to the voiced section to be identified.
The identification unit 44 recognizes the interval from the start of the voiced segment detected by the voiced segment start time detection unit 42 to the end of the voiced segment detected by the voiced segment end time detection unit 12. A section is defined, and an input signal in the recognition section is identified with reference to the reference parameter 22 of the voiced hidden Markov model stored in the storage unit 46.

Here, the hidden Markov model is a model in which speech is described by a statistical method, and the transition probability
It is composed of parameters such as output probability. As shown in FIG. 5, a sound section refers to a section in which a speech to be identified exists in an input signal, and a silent section refers to a section other than a speech section that is irrelevant to the identification target.
The voiced hidden Markov model refers to a hidden Markov model that has been learned in advance for each identification object using a signal in a voiced section of the learning data, and is prepared for each syllable and stored in the storage unit 46. Have been.

[0004]

By the way, the end point of a sound section can be detected by using a method described in, for example, JP-A-62-237498. When recognizing voice, some unvoiced consonants are difficult to distinguish from noises such as breathing sounds and lip noises. There is a problem that the detection at the start of the sound section is erroneous. Such an erroneous detection at the start of the sound section may cause erroneous recognition when the identification unit 44 performs the identification processing.

[0005] In order to solve such a problem, a speech recognition apparatus 50 having a configuration shown in FIG. 6 has been proposed. The voice recognition device 50 includes a storage unit 54 and an identification unit 5
2 is provided. In the storage unit 54, in addition to the reference parameters 22 of the voiced Hidden Markov Model corresponding to the voiced part of each identification target, the reference parameters 24 of the silent Hidden Markov Model corresponding to the silent section, and the Hidden Markov Model Network 56 Are stored. Here, the silent Hidden Markov Model refers to a Hidden Markov Model that has been learned in advance using a signal in a silent section of the learning data. Also,
The hidden Markov model network 56 is an example of a case where a single syllable is identified. In this example, a voiced hidden Markov model is prepared for each syllable, and a silent hidden Markov model + a voiced hidden Markov model + silent. Hidden Markov models are combined in order. Further, the identification unit 52
Calculates a known Viterbi score based on the hidden Markov model network 56 while referring to the reference parameter 22 of each sounded Hidden Markov model and the reference parameter 24 of a silent Hidden Markov Model, and calculates the calculated Viterbi score. Is detected, and the generated content corresponding to the voiced hidden Markov model on the detected path is output to the display device 32 or the like as a recognition result. In other words, the speech recognition apparatus 50 shown in FIG. 6 attempts to solve the above-described problem caused by the detection error at the start of the sound section by recognizing the occurrence without detecting the sound section. It is.

As shown in FIG. 7, the input signal may include a signal due to a breathing sound or a lip noise after a sound section. Therefore, when the speech recognition of such an input signal is performed by using the speech recognition apparatus 50 shown in FIG. 6 described above, when detecting a network path that maximizes the Viterbi score, a silent Markov model is The transition point to the Hidden Markov Model may be the point indicated by P in FIG. This is equivalent to performing the identification process on the section including the lip noise and the respiratory sound using the sound hidden Markov model, which means that the true sound section is not identified. That is,
The speech recognition device 50 shown in FIG. 6 has a problem that erroneous recognition may occur because no identification is performed for a true voiced section. As described above, Japanese Patent Application Laid-Open No. 62-237
The method described in Japanese Patent Publication No. 498 and the speech recognition apparatus 50 shown in FIG. 6 have a problem that the recognition accuracy is low.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a speech recognition device capable of improving recognition accuracy and a storage medium storing a computer program for performing speech recognition by the speech recognition device.

[0008]

According to the present invention, in order to achieve the above object, according to the first aspect of the present invention, a silent reference parameter corresponding to a signal in a silent section of a voice signal, and an audio signal reference Storage means for storing a sound reference parameter corresponding to a signal in a sound section; a sound section end time detecting means for detecting an end time of a sound section of the input signal as a sound section end time; Recognition start time setting means for setting a predetermined time in a range after the start time of the signal and before the start time of the sound section as a recognition start time for starting voice recognition, and the recognition start time setting means For the input signal from the set recognition start point to the end of the sound interval detected by the sound interval end detection means, the silent reference parameter stored in the storage means and With reference to the sound reference parameter, and identifying means for identifying the utterance contents of the input signal by the time normalization recognition method employs a technical means that are provided.

According to a second aspect of the present invention, in the speech recognition apparatus according to the first aspect, a technical means is adopted in which the time-normalized identification method is a hidden Markov model method.

Further, according to the third aspect of the present invention, a silent reference parameter created by learning from a signal in a silent section of an audio signal and a signal created by learning a signal in a voiced section of the audio signal are created. And the end point of the sound section of the input signal is detected as the end point of the sound section, and after the start point of the input signal, and the start of the sound section. A predetermined time within the range before the time is set as a recognition start time at which to start voice recognition, and from the set recognition start time to a sound section end time detected by the sound section end time detection means. A method for identifying an utterance content of the input signal by a time-normalized identification method while referring to a silent reference parameter and a sound reference parameter stored in the storage unit for the input signal. Adopt the technical means of the computer program is stored comprising a computer program.

[0011]

According to the first and second aspects of the present invention, the storage means stores a silent reference parameter corresponding to a signal in a silent section of the audio signal and a signal corresponding to a signal in a voiced section of the audio signal. A voiced reference parameter, the voiced segment end time point detection means detects an end time point of a voiced section of the input signal as a voiced segment end time point, and the recognition start time point setting means detects the input signal. A predetermined time after the start time and within a range before the start time of the sound section is set as a recognition start time at which the speech recognition is started, and the identification unit performs the recognition set by the recognition start time setting unit. For the input signal from the start time to the end of the sound period detected by the sound period end time detection means, the silent reference parameter and the sound reference parameter stored in the storage means are stored. With reference to identify the utterance contents of the input signal by a time normalized identification techniques. That is, the recognition start time setting means is configured to set the recognition start time to a predetermined time within the range after the start time of the input signal and before the start time of the sound section. Since the start time can be set to any time within the above range, it is not necessary to make the recognition start time coincide with the sound segment start time. Therefore, the conventional
Recognition is started due to erroneous determination of noise such as lip noise and unvoiced consonants, as in a method of performing recognition with the start of a voiced section as the start of recognition (Japanese Patent Laid-Open No. 62-237498). The time point does not coincide with the start point of the sound section, and there is no problem of incorrect recognition.

[0012] Further, the discriminating means is provided for the input signal from the recognition start time set by the recognition start time setting means to the voiced section end time detected by the voiced section end time detecting means. By referring to the silent reference parameter and the sound reference parameter stored in the storage unit, the utterance content of the input signal is identified by the time normalization identification method, so that noise such as lip noise is distinguished from asexual consonants. be able to. In other words, according to the recognition start time setting means, since the recognition is started before the sound section start time, it is possible that the recognition range includes noise such as lip noise. Those noises can be identified as silence. Therefore, there is no means for detecting the start point of the sound section and the end point of the sound section.
As in the case of the conventional speech recognition apparatus 50 (FIG. 6) that detects and recognizes a network path that maximizes a score, it is unlikely to cause erroneous recognition due to respiratory sounds, lip noise, and the like. Can be.

[0013] In particular, in the invention according to the second aspect, the time normalization identification method employs a technical means that is a hidden Markov model method, that is, uses a statistical processing method. Thus, it is possible to realize a speech recognition device having higher identification accuracy than the case of using the generalized identification method.

Further, in the invention according to claim 3, a silent reference parameter created by learning from a signal in a silent section of the audio signal and a signal created by learning a signal in a voiced section of the audio signal are created. And the end point of the sound section of the input signal is detected as the end point of the sound section, and after the start point of the input signal, and the start of the sound section. A predetermined time point within the range before the time point is set as a recognition start time point for starting voice recognition, and from the set recognition start time point to a sound section end time point detected by the sound section end time point detecting means. A method for identifying the utterance content of the input signal by a time-normalized identification method while referring to the silent reference parameter and the sound reference parameter stored in the storage means for the input signal. Since the computer program comprising computer program is configured that the stored storage medium, by using the storage medium, it is possible to realize a speech recognition apparatus according to claim 1. That is, the speech recognition device includes, for example, a CPU built in the speech recognition device,
Alternatively, since the computer is controlled by a computer connected to the speech recognition device, a storage unit as the storage medium is provided in the speech recognition device, or a computer program stored in the storage medium is installed in the computer. This is because the invention described in claim 1 can be implemented.

[0015]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the speech recognition apparatus according to the present invention will be described below with reference to FIGS.
FIG. 1 is an explanatory diagram showing a schematic configuration of the speech recognition apparatus of the present embodiment by blocks, FIG. 2 is an explanatory diagram showing a recognition start time point and a sound section end time point, and FIG. 5 is a flowchart showing a flow of processing of the voice recognition device of FIG. The same components as those in the related art are denoted by the same reference numerals, and description thereof is omitted.

First, the main configuration of the audio device according to the present embodiment will be described with reference to FIGS. As shown in FIG. 1, the speech recognition apparatus according to the present embodiment includes a speech section end point detection unit 12 constituting a speech section end point detection unit of the present invention, a storage unit 20 constituting a storage unit, and a recognition unit. The apparatus includes a recognition start time setting unit 14 constituting a start time setting unit and an identification unit 16 constituting an identification unit. The sound section end point detection unit 12 inputs a signal input to the voice input device 30 and digitized and output as an input signal,
The end point of the sound period shown in FIG. 2 is detected using the power of the input signal. Here, as described in the section of the related art, the voiced section refers to a section where the voice to be identified exists in the input signal, and the silent section refers to a section irrelevant to the identification target other than the voiced section. Point.

The storage unit 20 stores a parameter 22 of a voiced hidden Markov model corresponding to a voiced portion of each identification target, a parameter 24 of a voiceless hidden Markov model corresponding to a voiceless section, and a hidden Markov model network 26. I remember. Parameter 22 of the voiced hidden Markov model
Constitutes the voiced reference parameter of the present invention, and the parameter 24 of the silent hidden Markov model constitutes the silent reference parameter. Here, the hidden Markov model is a known model in which speech is described by a statistical technique, and is configured by parameters such as a transition probability and an output probability.

The voiced hidden Markov model refers to a Markov model that has been learned in advance for each identification object using a signal in a voiced section of the learning data. In the present embodiment, a hidden Markov model for each syllable is used. Are stored in the storage unit 26. Further, the silent Hidden Markov Model refers to a hidden Markov model that has been learned in advance using a signal in a silent section of the learning data. The hidden Markov model network 26 shown in FIG. 1 is an example of a case where a single syllable is recognized. In the present embodiment, a voiced hidden Markov model is prepared for each syllable, and a silent hidden Markov model +
These are combined in the order of the voiced hidden Markov model.
The recognition start time setting unit 14 sets an appropriate time in the section from the time after the start of the input signal to the time before the start of the sound section as the recognition start time. In the present embodiment, since the hidden Markov model network is used, it is not necessary to make the recognition start time coincide with the sound section start time, and it is sufficient if there is a sound section start time after the recognition start time, as shown in FIG. As described above, the time a that is sufficiently longer than the sound period of a single syllable from the sound period end time t3 detected by the sound period end detection unit 12, for example, about 500 m in the present embodiment.
A time point t1 returned by the second is set as a recognition start time point.

The identification unit 16 stores the input signal from the recognition start time t1 set by the recognition start time setting unit 14 to the voiced segment end time t3 detected by the voiced segment end time detection unit 12, and stores it in the storage unit. The Viterbi score is calculated by the known Viterbi method while referring to the parameter 22 of the voiced hidden Markov model and the parameter 24 of the silent hidden Markov model based on the hidden Markov model network 26 stored in 20. V
A network path that maximizes the iterbi score is detected, and an identification target corresponding to a sounded hidden Markov model on the detected network path is output to the display device 32 or the like as a recognition result.

Next, the processing contents of the speech recognition apparatus 10 of the present embodiment will be described with reference to the flowchart of FIG. Note that the speech recognition device 10 of the present embodiment
The computer program for executing the processing shown in FIG. 3 is stored in the storage unit 20. First, the sound section end point detection unit 12 detects the sound section end point of the input signal using the power of the input signal or the like (step 2). The detection is performed, for example, as described in
This is performed using the technique described in US Pat. Subsequently, the recognition start time setting unit 14 sets a time point 500 msec back from the sound section end time point t3 detected in step 2 as the recognition start time point t1 (step 4).

Subsequently, the identification unit 16 applies a known linear signal to the input signal in the range from the recognition start time t1 set in step 4 to the sound segment end time t3 detected in step 2 above. The feature amount is extracted using prediction or a filter bank (step 6), the hidden Markov model network 26 stored in the storage unit 20 is read (step 8), and based on the read hidden Markov model network 26. Then, the Viterbi score of the hidden Markov model network 26 is maximized by the known Viterbi method while referring to the parameters 22 of the sounded hidden Markov model and the parameters 24 of the silent hidden Markov model stored in the storage unit 20. A network path is extracted (step 10), and the extracted network is extracted. And outputs the verification area corresponding to a sound hidden Markov model on work path as the recognition result (step 12). The output recognition result is displayed on the display device 32 or the like.

As described above, if the speech recognition apparatus 10 of the present embodiment is used, the recognition start time t1 is a predetermined time within the range after the start time of the input signal and before the sound segment start time t2. Can be set to a point in time. Therefore, unlike the conventional method (Japanese Patent Application Laid-Open No. 62-237498), erroneous recognition is not caused due to the inability to match the start point of the sound section with the start time of recognition. Also, based on the hidden Markov model network 26, the parameter 22 of the sounded hidden Markov model and the parameter 2 of the silent hidden Markov model
4, the Viterbi method of the hidden Markov model network 26 is performed by the known Viterbi method.
The conventional speech recognition apparatus 50 (FIG. 4) extracts a network path that maximizes a score and outputs an identification target corresponding to the sounded hidden Markov model on the extracted network path as a recognition result. As described above, the probability of erroneously recognizing respiratory sounds, lip noises, and the like as speech can be reduced, so that recognition accuracy can be improved.

In the above embodiment, the hidden Markov model and the sounded hidden Markov model are connected using the hidden Markov model network, but the silent hidden Markov model and the sounded hidden Markov model are directly connected. , Other processing methods can be used. Further, the voiced hidden Markov model may be a model in which hidden Markov models such as phonemes are directly connected to each other in a smaller unit, or a model in which they are connected by a hidden Markov model network. Further, in the above-described embodiment, the hidden Markov model method is used as the time-normalized identification method, but a dynamic programming method (DP matching method) may be used. Further, in the above-described embodiment, a device that performs monosyllabic recognition has been described as a typical example of the voice recognition device of the present invention. However, the present invention can also be applied to a voice recognition device that performs word recognition, sentence recognition, and the like. . Further, in the above-described embodiment, the breathing sound is processed as a silent section. However, the breathing sound is adjacent to the sounding section, and its influence is smaller than that of the lip noise.

In the above embodiment, the computer program is stored in the storage unit 20. However, the computer program is stored in a CD-ROM, a floppy disk, or the like, and the computer program is read by the device. The device can also be operated by installing using the device. In this case, the CD-RO
M, FD, etc. function as the storage medium according to claim 3. Furthermore, a computer program can be read from an external information processing device via a wired or wireless communication unit and operated.

[0025]

As described above, claims 1 and 2 are as described above.
According to the invention described in (1), a recognition start time setting means for setting a predetermined time within a range after a start time of an input signal and before a start time of a sound section as a recognition start time at which speech recognition is started. Is provided, it is possible to eliminate erroneous recognition caused by the fact that the start point of the sound segment does not match the recognition start point. Moreover, for the input signal from the recognition start time set by the recognition start time setting means to the end of the sound section detected by the sound section end time detection means, the silent reference parameter and the sound reference parameter are set. Since the identification means for identifying the utterance content of the upper force signal by the time-normalized identification method while referring is provided, the recognition accuracy is not affected by lip noise, breathing sound, and the like. That is, according to the first and second aspects of the present invention, it is possible to realize a speech recognition device capable of improving recognition accuracy.

In particular, according to the second aspect of the present invention, since the time-normalized identification method uses the hidden Markov model method, which is a statistical processing method, the time-normalized identification method is more distinguishable than when other time-normalized identification methods are used. A highly accurate speech recognition device can be realized.

Further, according to the third aspect of the present invention,
A silence reference parameter created by learning from a signal in a silence section of the audio signal and a speech reference parameter created by learning a signal in a speech section of the audio signal are stored, and The end point of the sound section of the input signal is detected as the end point of the sound section, and the voice recognition is started at a predetermined point in time after the start point of the input signal and before the start point of the sound section. The input signal from the set recognition start point to the end of the sound section detected by the sound section end point detection means is set to the silence stored in the storage means. A computer program including a computer program for identifying the utterance content of the input signal by a time-normalized identification method while referring to a reference parameter and a sound reference parameter. The storage medium is stored in the speech recognition device, or the computer program stored in the storage medium is connected to the speech recognition device or the speech recognition device. By installing the speech recognition device in a computer, the speech recognition device according to the first aspect can be realized.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram showing a schematic configuration of a speech recognition device according to an embodiment of the present invention by blocks.

FIG. 2 is an explanatory diagram showing a recognition start point and a sound section end point in the embodiment of the present invention;

FIG. 3 is a flowchart showing a processing flow of the voice recognition device according to the embodiment of the present invention.

FIG. 4 is an explanatory diagram showing a schematic configuration of a conventional voice recognition device in blocks.

FIG. 5 is an explanatory diagram showing a sound section and a silent section;

FIG. 6 is an explanatory diagram showing a schematic configuration of a conventional voice recognition device by blocks.

FIG. 7 is an explanatory diagram showing a recognition section of the speech recognition device shown in FIG. 6;

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 10 Speech recognition device 12 Voice section end time detection part (voice section end time detection means) 14 Recognition start time setting section (recognition start time setting means) 16 Identification section (identification means) 20 Storage section (storage means) 22 Yes Parameters of sound hidden Markov model (sound reference parameter) 24 Parameters of silent hidden Markov model (silence reference parameter) 26 Hidden Markov model network

Claims (3)

[Claims] A storage means for storing a silent reference parameter corresponding to a signal in a silent section of the audio signal and a voiced reference parameter corresponding to a signal in a silent section of the audio signal; A sound section end time detecting means for detecting an end time of the sound section as a sound section end time; and a predetermined time within a range after the start time of the input signal and before the start time of the sound section. Start time setting means for setting the start time of voice recognition as a recognition start time point; and the end of the sound section detected by the sound section end time point detection means from the recognition start time set by the recognition start time setting means. For the input signal up to the point in time, while referring to the silent reference parameter and the voiced reference parameter stored in the storage means, the time normalized identification A speech recognition device, comprising: identification means for identifying utterance content; 2. The speech recognition apparatus according to claim 1, wherein the time-normalized identification method is a hidden Markov model method. 3. A non-speech reference parameter created by learning from a signal in a non-speech section of the audio signal, and a non-speech reference parameter created by learning a signal in a non-speech section of the audio signal are stored. Further, the end point of the sound section of the input signal is detected as the end point of the sound section, and the predetermined time within the range after the start point of the input signal and before the start point of the sound section. Is set as the recognition start time at which the voice recognition is started, and the storage means is used for the input signal from the set recognition start time to the sound section end time detected by the sound section end time detection means. A computer program for identifying the utterance content of the input signal by a time-normalized identification method while referring to the silent reference parameter and the sound reference parameter stored in the computer. A storage medium storing a computer program.