JPH0950288A - Device and method for recognizing voice - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accelerate the voice recognition speed by word spotting. SOLUTION: A voice signal continuously inputted to a voice input means 2 is converted into a voice pattern by a characteristic extraction means 3. A voice section predicted to include a recognition voice is detected by a voice detection means 4 from the voice pattern, and the voice recognition by the word spotting is executed by a voice recognition means 5 in the voice section. Since the word spotting is executed only in the voice section even though the voice signal is continuous, the processing load is reduced, and the recognition processing is speeded up.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device for recognizing continuous voice.

[0002]

2. Description of the Related Art At present, a speech recognition device for recognizing a speech uttered by a human has been developed. In such a voice recognition device, various kinds of information can be taken in without requiring manual operation of a keyboard, so that desired information can be input even in a work environment in which a person uses both hands, for example. In a personal computer that realizes such a voice recognition device, when a human generates a specific voice, it recognizes the specific voice and executes a predetermined processing operation.

When a human utters only one word, it is not difficult for the speech recognizer to recognize it, but in a natural human conversation, the voice is continuous, and in addition to the word It includes sighing and breathing. In the case of extracting and recognizing a word from continuous speech in this way, conventionally, continuous speech is fetched from the beginning to the end, and the speech recognition is executed as a whole. However, in this case, unnecessary voices such as sneezing and coughing become noises, the time required for voice recognition unnecessarily increases, and the recognition accuracy also deteriorates.

In word spotting, which is proposed as one of the methods for solving such a problem, only a portion matching a standard pattern is recognized from a voice signal. In other words, since there is no problem in recognizing the voice signal, it is possible to eliminate the adverse effects of unnecessary voice such as stagnation.

[0005]

In the word spotting as described above, only the necessary voice can be recognized even if the unnecessary voice exists in the continuous voice signal.

However, even in such word spotting, since the entire continuous voice signal is processed,
The processing load is large and it is difficult to execute the processing at high speed. Also, in such word spotting, unexpected erroneous recognition may occur unless the start and end of the recognition process are confirmed at proper timing.

Further, as described above, it has been put into practical use to operate a personal computer by using a voice recognition device, but it is expected that it may be difficult to quickly operate such equipment by voice. It For example, even if the device is set to stop various operations when it recognizes the voice "Stop," a person who finds a malfunction is "Ah ...
It's weird ... Stop! ”, The recognition process of“ Stop ”is executed after the recognition process of“ Ahhh… weird… ”is completed, and this recognition is delayed and the device cannot be stopped quickly.

[0008]

According to a first aspect of the present invention, there is provided a voice recognition device, wherein a voice input means for continuously inputting a voice signal and a continuous voice signal are converted into a voice pattern which is a time series of feature vectors. Feature extracting means, a voice detecting means for detecting a voiced section expected to include a voice to be recognized from a voice pattern according to a predetermined condition, and a voice recognition means for executing voice recognition by word spotting in the voiced section. Have and. Therefore, word spotting is not executed in a silent portion even if the audio signal is continuous. The voiced section in the present invention is a section in which a voice actually exists in a continuous voice signal, and is detected as, for example, a portion where the voice power is equal to or higher than a threshold.

In the voice recognition device according to the second aspect, the section extending means for extending the voiced section to at least one of the front and rear is provided. Therefore, the loss of voice due to a detection error is eliminated.

According to a third aspect of the present invention, there is provided a voice recognition device, which is provided with a silence detecting means for detecting a silent segment from a voice segment for a predetermined time or more, and segment dividing means for dividing the voice segment into a plurality of voice segments. The recognition means performs word spotting on each of the divided voiced sections. Therefore, the silent section is excluded even within the sound section. In addition,
The voiced section referred to in the present invention is a section in which no voice actually exists in a continuous voice signal, and is detected as, for example, a portion where the voice power is below a threshold value.

According to a fourth aspect of the present invention, there is provided a voice recognizing device, which is provided with a power detecting means for detecting the power of each of a plurality of voiced sections divided by a silent section, and the voice recognizing means has a maximum voiced section. Perform word spotting first. Therefore, a voice uttered by a human being is first recognized.

According to a fifth aspect of the present invention, there is provided a voice recognition device, wherein power detection means is provided for detecting the power of each of a plurality of voiced sections divided by a voiceless section, and the voice recognition means has a voice output whose power exceeds a threshold value. Perform word spotting preferentially in sections. For this reason, the voice uttered by a human being is loudly recognized with priority.

According to a sixth aspect of the present invention, the voice recognition method includes converting a voice signal continuously input to the voice input means into a voice pattern which is a time series of feature vectors by the feature extraction means, and includes recognized voice. The voiced section that is expected to be detected is detected from the voice pattern by the voiced detection unit according to a predetermined condition, and voice recognition by word spotting is executed by the voice recognition unit in this voiced section. Therefore, word spotting is not executed in a silent portion even if the audio signal is continuous.

[0014]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below with reference to the drawings. First, as shown in FIG. 1, the voice recognition device 1 illustrated here has a voice input unit 2 which is a voice input unit, and the voice input unit 2 has a feature extraction unit which is a feature extraction unit. 3, a section detecting unit 4 which is a voice detecting unit, a voice recognizing unit 5 which is a voice recognizing unit, and a result output unit 6 which is a result outputting unit are sequentially connected.

The voice input unit 2 includes a microphone or an A / DC (Analog / Digital Converto) as hardware.
r) etc., and converts continuous sound into digital signals. The feature extraction unit 3 has a microcomputer and converts a voice signal into a voice pattern which is a time series of feature vectors.

Although various existing methods can be used for converting a continuous voice into a digital signal and then converting into a voice pattern as described above, here, the continuous voice at 16 (kHz) is 16
Convert to (bit) digital signal. This is a tenth order LPC (Linear Pred
ictive Coding) cepstrum, the feature vector of continuous speech is generated as a 10-dimensional vector for every 10 (ms) frame.

The section detecting section 4 has a microcomputer and detects a sound section in which the recognized voice is expected to be included from the voice pattern according to a predetermined condition. This condition is set based on the comparison between the voice power and the threshold value, and more specifically, when the voice power becomes equal to or higher than the threshold value, the start of the voiced section is detected. If the state in which the voice power is equal to or lower than the threshold continues for a predetermined time after the detection of the start of such a section, the time when the voice power becomes equal to or lower than the threshold is detected as the end of the voiced section. Such processing operations are executed in frame units and managed by frame numbers.

The voice recognition unit 5 has a microcomputer and executes voice recognition by word spotting in a voiced section. The result output unit 6 has an interface or the like, and is connected to, for example, a device operated by voice. In addition, each part 2-6 mentioned above
A control circuit is connected to the control circuit, and the control circuit is provided with a start switch.

With such a configuration, the above-described voice recognition device 1 extracts and recognizes words from the voices that a human continuously utters. The voice recognition method of the voice recognition apparatus 1 will be described in detail below with reference to FIGS. 2 to 4.

First, as shown in FIGS. 2 and 4, when the start switch is operated, the control circuit activates the respective units 2 to 6, and the voice input unit 2 converts the voice continuously input from the outside. , The continuous voice signal is converted into an LPC cepstrum by the feature extraction unit 3.

As shown in FIG. 3, the section detector 4 compares the voice power with the threshold Pt and detects the first frame in which the voice power is equal to or higher than the threshold Pt. This is detected as the start frame of the voiced section, and its frame number Is is recorded. When this is completed, the voice power and the threshold value Pt are compared, and when the voice power becomes equal to or lower than the threshold value Pt, the frame number is temporarily stored as the end point candidate of the voiced section. In this way, it is determined whether or not the state in which the voice power is equal to or lower than the threshold Pt continues for a predetermined time Lt. When this continuity is detected, the end of the voiced section is detected, and the temporarily stored frame number Ie is determined. .

When the voiced section "Is to Ie" is thus detected, the voice recognition section 5 executes the voice recognition by word spotting only in the voiced section, and the recognition result is output from the result output section 6. To be done.

In the voice recognition device 1 described above, even if the start switch is used to input the voice recognition, the voice recognition by word spotting is executed only in the voiced section in which the recognized voice is expected to be included. Since the entire continuous audio signal is not processed, the processing load can be reduced and the recognition processing can be sped up, and the start and end of the recognition processing can be determined at appropriate timings. False recognition is also prevented from occurring.

The present invention is not limited to the above-described embodiment, but allows various modifications. For example, in the voice recognition device 1 described above, the voice power is compared with the threshold value to detect the voiced section, but the section extension means for extending such a voiced section to at least one of the front and back is provided. Is also possible.

In this case, since the voiced section is extended forward and backward, the loss of voice due to a detection error is eliminated and the accuracy of voice recognition is improved. For example, the start frame “I
By subtracting a predetermined number of frames "L1" from "s", the voiced section can be extended forward, and by adding a predetermined number of frames "L2" to the end frame "Ie" of the voiced section It is possible to extend the sound section backwards, such an extension being about 50 (ms) = 5 frames, which is the length of the sound in the front, and "100 (ms), which is the sounding length, in the back. ) = About 10 frames ”is preferable.

Further, in the above-described voice recognition device 1, the whole of the voiced section detected from the continuous voice signal is exemplified as the processing target of the word spotting.
As shown in FIG. 3, a silence detecting unit that detects a silent period of “L3” or more for a predetermined time from the voice period “Is to Ie” and a segment dividing unit that divides the voice period into a plurality of voice periods are provided. It is also possible to execute word spotting in each of the divided voiced sections.

In this case, since the silent section is eliminated even within the sound section, the processing load can be further reduced and the recognition processing can be speeded up. The start and end of the recognition processing can be optimized and the recognition accuracy can be improved. Can be improved. If voice recognition is executed in units of words or syllables, pauses and breaths are silent intervals that should be excluded, and consonants are silent intervals that should not be excluded. In such a case, the predetermined time for detecting the silent section may be set to a time longer than the consonant and shorter than a breathing, so this is about “300 (ms) = 30 frames”.

When one voiced section is divided into a plurality of voiceless sections in this way, in addition to detecting one voiced section detected as described above and then dividing it into a plurality of sections,
It is also possible to detect a plurality of divided voiced sections in order from the beginning.

Furthermore, it can be assumed that an external device is connected to the result output unit 6 of the voice recognition device 1 for dividing the voiced section into a plurality of parts in this way, and the external device is operated according to the recognition result of the voice recognition device 1. . In such a case, it is preferable to provide a power detection means for detecting the power of each of the voiced sections divided by the voiceless section, and execute the word spotting first in the voiced section having the maximum power.

For example, if an electronic file device is connected to the voice recognition device 1 and various operations thereof are voice-controlled, it is set so that the various operations are stopped by a "stop" voice. In such a system, the person who discovered the malfunction of the electronic file device said, "Oh ... it's weird ...
stop! When you say "", this "stop" voice is naturally made loud. In this case, the voice recognition device 1
Detects multiple voiced sections in the order of "Are", "Weird" and "Stop", but "Stop" with the maximum power
Is recognized first, the electronic file device can be stopped quickly.

Similarly, it is possible to detect the power of each of a plurality of voiced sections obtained by dividing a silent section and preferentially execute word spotting in a voiced section in which the power exceeds a threshold value. In this case, the voice recognition device 1 recognizes continuous voices in the order of loudness of human utterances. Therefore, useful voices can be recognized quickly even if louder and unnecessary voices are uttered than useful voices. it can.

[0032]

According to the voice recognition apparatus of the present invention, the voice input means for continuously inputting the voice signal and the feature extracting means for converting the continuous voice signal into a voice pattern which is a time series of feature vectors. And voiced means for detecting a voiced section expected to include a recognized voice from a voice pattern according to a predetermined condition, and voice recognition means for executing voice recognition by word spotting in the voiced section. Thus, even if a voice signal is continuously input, voice recognition by word spotting is executed only in a voiced section where the recognized voice is expected to be included, and the entire continuous voice signal is not processed. Therefore, the processing load can be reduced and the recognition processing can be speeded up.

In the voice recognition apparatus according to the second aspect, since the section extending means for extending the voiced section to at least one of the front and back is provided, it is possible to prevent the voice from being lost due to the detection error of the voiced section. The recognition accuracy of can be improved.

In the voice recognition apparatus according to the third aspect of the invention, there is provided silence detecting means for detecting a silent section of a predetermined duration or more from the sound section, and section dividing means for dividing the sound section into a plurality of silent sections. By performing word spotting on each of the plurality of divided voiced sections, the recognition means can eliminate a silent section that is not the target of word spotting from the voiced section, further reducing the processing load. The recognition process can be speeded up.

In the voice recognition apparatus according to the present invention, a power detecting means for detecting the power of each of the plurality of voiced sections divided by the silent section is provided, and the voice recognition means detects the voiced section having the maximum power. By first performing word spotting, for example, when a system is created in which various actions are stopped by the voice "stop", the person who finds the malfunction says "Are ... strange ... stop!" Even so, the system will be able to stop quickly because the naturally pronounced "stop" is recognized first.

According to a fifth aspect of the present invention, there is provided a voice recognizing device, which is provided with a power detecting means for detecting the power of each of the plurality of voiced sections divided by the voiceless section, and the voice recognizing means has the voice detecting section whose power exceeds a threshold value. In the case of forming a system in which various operations are stopped by the voice of "stop" by preferentially executing word spotting in a section,
Even if the person who found the malfunction utters, "Ahhh ... weird ... stop!", It naturally recognizes a loud "stop", so the system can be stopped quickly. Even if a louder scream is made, the next loud “stop” is quickly recognized.

In the voice recognition method according to the sixth aspect, the voice signal continuously input to the voice input means is converted into the voice pattern which is the time series of the feature vector by the feature extraction means, and the recognized voice is included. It is possible to continuously input a voice signal by detecting a voiced section expected from the voice pattern by a voiced detection unit according to a predetermined condition and executing voice recognition by word spotting in this voiced section by the voice recognition unit. However, the voice recognition by word spotting is executed only in the voiced section where the recognized voice is expected to be included, and the entire continuous voice signal is not processed, so the processing load is reduced. The recognition process can be speeded up.

[Brief description of drawings]

FIG. 1 is a block diagram showing a voice recognition device according to an embodiment of the present invention.

FIG. 2 is a flowchart showing a voice recognition method.

FIG. 3 is a flowchart showing a processing operation of section detection.

FIG. 4 is a time chart showing an LPC cepstrum that is a voice pattern.

FIG. 5 is a time chart showing a modified LPC cepstrum.

[Explanation of symbols]

 1 voice recognition device 2 voice input means 3 feature extraction means 4 voiced detection means 5 voice recognition means

Claims (6)

[Claims] 1. A voice input means for continuously inputting a voice signal, a feature extracting means for converting the continuous voice signal into a voice pattern which is a time series of feature vectors, and a voice to be recognized. A voice recognition device comprising: a voiced voice detection unit that detects an expected voiced segment from a voice pattern in accordance with a predetermined condition; and a voice recognition unit that performs voice recognition by word spotting in the voiced segment. 2. The voice recognition device according to claim 1, further comprising a section extension means for extending a voiced section to at least one of front and rear. 3. A silence detecting means for detecting a silent interval of a predetermined time or more from a voice interval is provided, and a segment dividing means for dividing the voice interval into a plurality of voice intervals is provided, and the voice recognizing means divides word spotting. The voice recognition device according to claim 1, wherein the voice recognition device is executed in each of the plurality of voiced sections. 4. A power detection means for detecting the power of each of the voiced sections divided by the voiceless section is provided, and the voice recognition means first executes word spotting in the voiced section having the maximum power. The voice recognition device according to claim 4, wherein 5. A power detection means for detecting the power of each of the plurality of voiced sections divided by the voiceless section is provided, and the voice recognition means preferentially performs word spotting in the voiced section whose power exceeds a threshold value. The speech recognition apparatus according to claim 4, which is executed. 6. A voiced section in which a voice to be recognized is expected to be converted by converting the voice signal continuously input to the voice input means into a voice pattern which is a time series of feature vectors by the feature extraction means. The voice recognition method is characterized in that the voice recognition is detected from the voice pattern according to a predetermined condition and the voice recognition by word spotting is executed by the voice recognition means in the voiced section.