JP3125928B2 - Voice recognition device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a speech recognition device.

2. Description of the Related Art Speech recognition devices have been actively researched, and there are methods for specific speakers and unspecified speakers. In any case, it is an essential condition for correct recognition that a speech section can be cut out correctly. A method of extracting a voice section is known from Japanese Patent Publication No. 62-50837. This is because how to determine the threshold value A is important. If it is set low, the energy of the signal exceeds the threshold value due to noise, and a voice detection signal is output by the voice section cutout unit. On the other hand, if the threshold value A is set high, the beginning of the sound will be lost.
In order to prevent this, for example, JP-A-57-177197 and
Japanese Patent Application Laid-Open No. 58-076899, which shows how to set the threshold value A according to the level of ambient noise, determines the threshold value A in which the above-described inconvenience is unlikely to occur. However, the noise level is constantly changing, and in some cases, even if the noise level does not change, if the loudness of the voice of the speaker decreases, accurate cutout cannot be performed.

As another method, as disclosed in Japanese Patent Application Laid-Open No. 57-148799, there is a method in which not only speech energy but also a phoneme sequence is analyzed to cut out an accurate section, but a troublesome operation of phoneme classification is performed. There must be. Further, JP-A-56-5658
As disclosed in Japanese Patent Laid-Open Publication No. 8 (1999) -1995, there is a method of detecting the beginning of a speech section by the above method and returning the speech section 50 msec before the starting point. This compensates for the portion that is likely to be lost when the voice is small and the threshold is high, but the voice is not always included in this 50 ms. There is a drawback that the recognition accuracy is lowered when no voice is included or when noise is present.

FIG. 10 shows a method for compensating for the drawback when the threshold value A is used (Japanese Patent Laid-Open No. 60-23899). This is to integrate the energy in the voice section and normalize the energy by the time length of the voice section if the result is larger than a certain value, and instruct the user to utter a loud voice if the result is smaller. However, words with small parts of energy as shown in FIG.
In other words, there is a problem that it is difficult to treat the words as shown in FIG. For example, in FIG. 11 (a), since words such as "stop" include a prompting sound, there is an instruction to increase the voice while speaking in a loud voice, while words such as "eyes" In the type shown in Fig. 11 (b), even if the voice is low and the consonant / m / cannot be detected, the vowel / e / has a larger energy than / m /, so the average of the voice section is There is a drawback that the detection error often occurs because the threshold value A is often not lowered.

Object The present invention has been made in view of the above-mentioned circumstances, and is intended to accurately detect the beginning of a voice irrespective of the loudness of a speaker's voice, and to detect a correct voice section. As a result, the object of the present invention is to provide a speech recognition device with high recognition accuracy.

Configuration In order to achieve the above object, the present invention provides an audio-to-electric converter for converting a sound into an electric signal, a sound section detecting unit for extracting a portion related to the sound from the converted signal, A voice recognition device having a recognition unit that recognizes voice using a signal, analyzing an electrical signal that is continuously present temporally before the candidate of the starting point of the voice detected by the voice section detection unit, and the analysis result In this case, the user is instructed how to utter in response to the request. Hereinafter, a description will be given based on examples of the present invention.

First, FIG. 5 is a block diagram of a voice recognition device using ordinary pattern matching. The signal from the microphone 12 is detected by the voice section detection unit 13 and the voice pattern from the microphone is compared with the standard pattern. 15 is recognized by the recognition unit 14.

FIG. 1 is a configuration diagram for explaining an embodiment of a voice recognition device according to the present invention. FIG. 1 shows the configuration of a voice section detection unit in FIG. , 2
Is an A / D converter, 3 is a first memory, 4 is an energy detector, 5 is a register, 6 is a comparator, 7 is a threshold A, 8 is an energy detector, 9 is a comparator, 10 is a threshold B, Reference numeral 11 denotes a connecting portion. A speech having a speech section detection unit for extracting a portion related to speech from a signal converted by an acoustic-to-electric converter that first converts speech into an electric signal, and a recognition unit for recognizing speech using the detected signal. A description will be given below of how the recognition device shifts the start point of the voice to a point before the start point of the voice section is detected.

The signal from the microphone 1 is A / D converted by the A / D converter 2 and written in the first memory 3 in order. At this time, the data may be converted into a feature value before writing, or the written data may be read and converted. The feature amount referred to here indicates an analysis result such as a spectrum or LPC, and the type is not particularly limited. The first memory 3 is written in such a manner that the data is sequentially shifted with the passage of time and returns to the top when the end is full. At the same time as writing to the memory, the energy of the signal is detected to detect a voice section.

In general, a method for detecting a voice section is to separate the voice section from surrounding background noise based on the magnitude of voice energy as shown in FIG. In this method, an energy threshold value A of the noise level is determined before a voice is input, and a portion from a point in time when a sound larger than the threshold value A is input to a point below the threshold value A is defined as a voice section. This is the basis of the idea, but various improvements have been made to distinguish it from noise. In addition, it is not necessary to use a specific feature amount, and any of the most general power spectrum, LPC, and cepstrum may be used. Taking the power spectrum as an example,
This can be realized by applying the input voice to the band-pass filter group, and the analysis method can be freely changed by selecting the characteristics of the band-pass filter.

Next, a description will be given with reference to the waveforms of FIG. When the illustrated audio waveform is input, the energy is calculated first, and the detected audio is stored in the register assuming that the audio section has started when the energy is larger than the threshold value A.
Assuming that data for 100 msec is stored in the first memory, the data has the data from a to d in FIG. Therefore, the starting point of the sound corresponding to c is shifted from a to c. However, when moving to a, the extra space between a and b is attached to the beginning of the sound. In particular, if this 100 ms is added to the energy waveform like the vowel in FIG.
In 100msec, it becomes almost unnecessary data. Therefore, after shifting the starting point, the signal between the starting point (previous starting point) detected by the voice section detection unit and the starting point (later starting point) formed by shifting is analyzed, and the subsequent starting point is determined according to the analysis result. Was moved. The embodiment of FIG. 1 focuses on the energies a to c as examples of this analysis. The energy of the data of a to c held in the first memory 3 is detected and compared with the threshold B. Needless to say, the threshold value B must be greater than the threshold value A> the threshold value B. The threshold value B may be zero. When the comparator 9 moves the beginning of the voice to the point where the threshold value B is exceeded, the correct voice section from b to d can be detected in FIG. 7 and the waveform as shown in FIG. Can be detected correctly. Of course, this method may be applied to the end of voice. Also, the buffer of 100 ms is not limited and may be shorter. In FIG. 7, by combining the parts b to c detected in this way with the parts c to d detected by the usual method,
The correct voice of d is obtained. This is transferred to the recognition unit for recognition. The recognition method is not particularly limited, and a known method such as DP matching may be used. In addition, FIG. 1 shows the energy detection unit divided into two parts. However, it is possible to use one part for both, and the threshold value does not have both values. For example, B = A / You may decide like 5. Furthermore, although the energy is shown as a method of analyzing the data of a to c here, the method can be executed by using another method such as obtaining a difference between power spectra.

Next, another embodiment of the present invention will be described with reference to FIG. In the figure, 16 is the second memory, 17 is the clear unit,
Other parts 1 to 11 which operate in the same manner as in FIG.
The same reference numerals as in FIG. 1 are used.

FIG. 2 is a diagram illustrating a speech section detection unit that extracts a portion related to speech from a signal converted by an acoustic-electric converter that converts speech into an electric signal, and a recognition unit that recognizes speech using the detected signal. In the speech recognition device having, when the start point is shifted a predetermined time before the start point of the voice detected by the voice section detection unit, the shifted section is analyzed, and when a silent section is detected in a portion other than the start point of the section, The starting point of the voice is moved to the end of the silent section. Microphone 1
The audio signal input from the CPU or a signal obtained by converting the audio signal into a feature value is recorded in the first memory 3. In the first memory 3, one or a plurality of data output at each timing can be sequentially recorded, and any memory that can temporarily store, for example, 100 msec of data may be used. Ten
When writing for 0 msec is completed, the next data is rewritten from the top to the top again. At the same time as writing to the first memory 3, the energy of each timing is obtained, and whether the energy is greater than the threshold value A is compared by the first comparator 6. Starts importing / D converted data. Next, the energy is obtained from the beginning of the data of 100 ms stored in the first memory 3 and compared with the threshold value B and the second comparator 9. Write to memory 16. Here, after that, if there is an energy portion smaller than the threshold value B, the entire contents of the second memory 16 are cleared by the clear function 17, and the same operation is repeated. After checking the data for 100 ms, if there is data in the second memory 16,
This is combined with the beginning of the data stored in the register 5 to generate audio data, which is transferred to the recognition unit. This will be described with reference to waveforms as shown in FIG. The first phoneme is missing in the voice section detected by the normal method. So the top
If data up to 100 ms before is taken, noise such as the sound of opening and closing the lips may be included together with the voice. Therefore, the energy for this 100 msec is checked again, and the rest is discarded except for the portion that is continuous with the voice section found earlier. Thereby, a correct voice section can be detected. In order to obtain this effect, the threshold value A must be larger than the threshold value B.

This method has been described at the beginning of speech, but can also be applied to the end of speech. Further, the memory of 100 ms is not limited, and may be further shortened. The recognition method is not particularly limited, and a known method such as the above-described DP matching may be used. Such technical content,
For example, it is described in detail in "Speech recognition" (by Niimi, Kyoritsu Shuppan).

In addition, FIG. 2 shows the energy detection unit divided into two for convenience as in the case of FIG. 1, but it is also possible to use one for both, and the threshold does not have both values. For example, B = A / 5 may be determined. Further, here, a to
Although energy was shown as a method of analyzing the data of c,
It can also be executed by using other methods such as taking a difference between power spectra.

Further, another embodiment according to the present invention will be described with reference to FIG. In the figure, reference numeral 18 denotes a display unit, and other parts 1 to 10 which operate in the same manner as in FIG. 1 are denoted by the same reference numerals as those in FIG. In this embodiment, attention is paid to the fact that the rising of the voice is not so steep, and particularly when the leading sound is a consonant, the sound gradually rises, so that this portion cannot be detected accurately. A speech recognition device including a speech section detection unit that extracts a portion related to speech from a signal converted by an acoustic-electric converter that converts speech into an electric signal, and a recognition unit that recognizes speech using the detected signal. In the above, data before the voice start point detected by the voice section detection unit is analyzed, and a method of uttering the user is instructed according to the analysis result.

Audio from the microphone 1 is converted to a digital signal by the A / D converter 2. In this case, it is desirable to perform A / D conversion after previously converting to the feature amount. This data is
Are sequentially recorded in the memory 3 for each sample time, and the energy is detected. This energy is compared with the threshold value A in the first comparator 6 and at a point above which it is considered as the start of the speech. In other words, this part detects the voice section shown in FIG.
When the energy becomes larger than the threshold value A, the data just before that is stored in the first memory 3. Assuming that 0.1 second of data can be recorded in the first memory 3, the data is recorded at the moment when the sound rises.
You will have data up to one second ago. Therefore, the energy for this 0.1 second is analyzed, and if it is larger than the determined threshold B, it is considered that what should be detected as a voice section is small in voice and energy is low, causing a detection error. Message is displayed on the display unit 18. If it is smaller than the threshold value B, it is determined that the rising is detected accurately, and nothing is displayed. Alternatively, a message indicating “good” is displayed. here,
The threshold value B is determined. When comparing the entire 0.1 seconds, the energy may be set to about 0.1 second of A / 2, and for comparison at each time point, the value may be about A / 2. Good.
By incorporating this into the voice section detection unit in FIG. 5, the recognition device of the present invention operates. The method of the recognition unit in this case is not particularly limited. In addition, a routine for registering the standard pattern in the specific speaker system is required. In FIG. 5, it is omitted considering unspecified speakers and the like.

Further, another embodiment according to the present invention will be described with reference to FIG. 19 is a feature extraction unit, 20 is a register group (register control unit), 21 is a content check unit, 22 is a ring counter, 23
Is a data transfer control unit, 24 is a start detection unit, 25 is a start correction unit, 26 is an input data buffer, 27 is a pattern matching unit, 28
Is a dictionary template, and 29 is a result output unit. First, the seventh
As shown in the figure, when determining the voice section, the feature amount of the voice,
For example, a part larger than the threshold A is detected as a voice section by comparing with a certain threshold A based on a power spectrum, an LPC cepstrum, or the like. However, if the threshold value A is constant with respect to the loudness of the voice and the ambient noise, it becomes difficult to detect the threshold value A. Therefore, in general, the threshold value A is often set variably. Therefore,
A consonant part at the beginning of a word cannot be detected due to the threshold value A when the voice is low or the ambient noise is high. Therefore, the influence is reduced by moving the starting end to the point b. or,
In a speech recognition device, generally, data is fetched until the end of voice utterance, and then collation is performed in parallel with data input (for example, DP matching,
Preliminary selection in the BTSP method), and it is necessary to detect voice sections in real time. However, it is difficult to retreat to the past in time, and in fact, past data is accumulated using a group of registers.

The feature amount of the unknown speech input at a certain sample period (for example, 10 ms, 20 ms) is obtained by the feature extraction unit 19, and the feature data is sequentially stored in the register group until the beginning of the speech is detected.
Stored in 20. The register group 20 has a predetermined length of n frames, and its input is set to 0, 1, 2,..., N, 0, 1, 2,. , 0,... Are controlled by the register control unit 20. On the other hand, when the start end is detected by the start end detection unit 24, the past n frames are similarly configured from the position of the m frame specified from the position indicated by the ring counter 22, so that the correction is performed simultaneously with the detection of the start end. It can be collated in real time.

Next, for the correction of the starting end, the value of the register group 20 indicated by the ring counter 22 is searched in the reverse direction, and the frame immediately before the 0 data becomes the true starting end. According to the pointer obtained in this way, the data from the true beginning to the current pointer is transferred to the input buffer, and then the data input according to the sample period is processed again in real time.

Advantages As is apparent from the above description, according to the present invention, it is possible to detect a correct voice section regardless of the volume of the voice of the speaker, so that the recognition rate of the voice recognition device can be improved, and the correction can be performed. Can be performed in real time.

In addition, the section can be detected accurately regardless of the level of the audio level. In particular, the present invention can realize an effective and high recognition accuracy for a person whose voice is small and easily utters the opening and closing sound of the lips.

Also, when speaking in a small voice, it is possible to instruct the user to accurately increase the subsequent utterance regardless of the uttered word. If the threshold value A fluctuates with the surrounding noise, the noise is loud, and when the threshold value A rises, an instruction is issued to produce a louder voice. As a result, the beginning and end of the voice can be accurately detected, and as a result, the recognition accuracy can be improved.

[Brief description of the drawings]

FIG. 1 is a configuration diagram for explaining one embodiment of a voice recognition device according to the present invention, FIG. 2 is a configuration diagram showing another embodiment, and FIG. 3 is a configuration showing still another embodiment. FIG. 4, FIG. 4 is a block diagram showing still another embodiment, FIG. 5 is a block diagram of a speech recognition device using ordinary pattern matching, and FIG. 6 is a diagram for explaining a method of detecting a speech section. , FIGS. 7 to 9 are diagrams showing threshold values and speech sections for speech waveforms, FIG. 10 is a diagram showing a conventional example of a speech segment detection method, and FIGS. 11 (a) and (b). FIG. 2 is a diagram showing a conventional example of a voice section detection method for voices having different energy distributions. 1 ... Microphone, 2 ... A / D converter, 3 ... First memory, 4,8 ... Energy detector, 5 ... Register, 6,9
...... Comparator 7, Threshold A, 10 Threshold B, 11 ...... Connection unit, 16 ... Second memory, 17 ... Clear, 18 ... Display unit.

Continuation of front page (56) References JP-A-62-191895 (JP, A) JP-A-63-77097 (JP, A) JP-A-60-23899 (JP, A) JP-A-60-101598 (JP, A) JP-A-59-152498 (JP, A) JP-A-58-76899 (JP, A) JP-B-62-50837 (JP, B2)

Claims (1)

(57) [Claims] An audio-to-electric converter for converting a sound into an electric signal, a sound section detecting unit for extracting a portion related to the sound from the converted signal, and recognizing the sound using the detected signal. In the voice recognition device having a recognition unit, the electrical signal that is continuously present temporally before the starting point candidate of the voice detected by the voice section detection unit is analyzed, and the utterance of the user is A speech recognition device characterized by instructing a way.