JP3430133B2 - Speech recognition device and program recording medium - Google Patents

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 a voice signal taken in and a program recording medium recording a voice data input conversion processing program.

[0002]

2. Description of the Related Art At present, a method using a keyboard, a mouse or the like is mainly used to instruct a predetermined operation to a personal computer or the like. However, on the above-mentioned personal computer, for example, when displaying a frequently viewed web page,
Start a browser and select a web page registered in "Favorites", or enter a complicated URL (uniform
It is necessary to enter the resource location), which is troublesome in any case. Therefore, voice recognition, which gives a command directly to a personal computer or the like by voice to perform a predetermined operation, has begun to attract attention.

Conventionally, as a voice recognition device of this type, FIG.
It is as outlined in. In this voice recognition device, an analog voice signal picked up by a microphone (not shown) is digitized by an A / D conversion circuit (not shown), and voice power calculated from the obtained digital voice signal is output. It is given to the section cutout unit 1. Then, the voice section cut out based on the voice power is sent to the acoustic analysis unit 2, and the feature vector of the voice section is extracted. The feature vector obtained in this way is sent to the matching unit 3, the matching unit 3 performs matching with the standard pattern registered in the dictionary unit 4 in advance, and the label of the most similar standard pattern is output as the recognition result. Of.

In this case, the A / D conversion circuit for converting an analog voice signal into a digital voice signal has a structure as shown in FIG. The low-pass filter 5 is an anti-alias filter, and the cutoff frequency may be variable. The amplifier 6 amplifies a weak analog signal to a size that the A / D converter 8 can easily handle, and in some cases the gain can be set by giving a digital signal of several bits. The sample hold 7 is for keeping the analog signal being converted in the A / D converter 8 constant, and needs to be controlled in close relation to the control of the A / D converter 8.

In the conventional speech recognition apparatus, the sampling frequency (12 kHz) twice the maximum frequency in the frequency band (for example, up to 6 kHz) processed by the speech recognition section composed of the acoustic analysis section 2 and the matching section 3 is used. The A / D conversion circuit is controlled to perform sampling to convert an analog voice signal into a digital voice signal.

[0006]

However, the above-mentioned conventional voice recognition device has the following problems.
That is, the frequency band (0 Hz
Since sampling is performed at a sampling frequency of 12 kHz, which is twice the maximum frequency in (~ 6 kHz), an audio signal in the frequency band (0 Hz to 6 kHz) is obtained as shown in FIG. In that case, as shown in FIG. 6, the cutoff characteristic of the anti-aliasing filter (low-pass filter 5) of the A / D conversion circuit in the speech recognition apparatus and the learning data used for learning the standard pattern or the acoustic model are used when the learning data is extracted. If the cut-off characteristic of the anti-aliasing filter is different, the frequency characteristic of the transition region is different, and the recognition rate is lowered.

When the cutoff characteristic of the antialiasing filter of the A / D conversion circuit in this speech recognition apparatus is known in advance, it is not possible to learn the standard pattern or acoustic model in consideration of the cutoff characteristic of the antialiasing filter. is there. However, if the cutoff characteristics of the anti-aliasing filter of the A / D conversion circuit installed in the sound card may differ depending on the model, such as in the case of a personal computer, consider the cutoff characteristics in advance and use a standard pattern or sound. The problem is that it is difficult to learn the model.

Therefore, an object of the present invention is to provide a voice recognition device capable of preventing a reduction in the recognition rate due to the cutoff characteristic of an anti-aliasing filter by a simple process, and a program recording medium recording a voice recognition processing program. .

[0009]

In order to achieve the above object, the first invention is a voice recognition device for recognizing an input voice by the voice recognition means, which is wider than the frequency band processed by the voice recognition means. A voice input unit having a pass band anti-aliasing filter, for sampling the input voice that has passed through the anti-aliasing filter to obtain voice data, and a sampling frequency of the voice input unit, wherein the voice recognition is performed in the pass band of the anti-aliasing filter. Input control means for setting the oversampling frequency so that the frequency band processed by the means is sufficiently included, and the same characteristics as the cutoff characteristics of the antialiasing filter used during learning data extraction for learning the standard pattern or acoustic model, or Use filters with similar characteristics Filtering the voice data sampled by the voice input unit, down-sampling the voice data after the filtering process, converting the voice data into voice data having a predetermined frequency band, and sending the voice data to the voice recognition unit. It is characterized in that it is provided with a data conversion means.

With the above arrangement, the voice input means samples the input voice at an oversampling frequency such that the pass band of the antialiasing filter sufficiently includes the frequency band processed by the voice recognition means. Then, the data conversion means performs the filtering process on the voice data by using a filter having the same characteristic or a similar characteristic to the cutoff characteristic of the antialiasing filter used at the time of extracting the learning data. In this way, the frequency characteristic of the input voice is matched with the frequency characteristic of the learning data for learning the standard pattern or the acoustic model, and the reduction of the recognition rate due to the cutoff characteristic of the antialiasing filter included in the voice input unit is prevented.

The speech recognition apparatus according to the first aspect of the invention is
It is desirable that the input control means is set to an oversampling frequency so that voice data having a frequency band n times (n: a positive integer of 2 or more) processed by the voice recognition means can be obtained.

According to the above arrangement, the 1 / n downsampling process by the data converting means is easily performed.

The speech recognition apparatus of the first invention is
It is desirable that the data conversion means has a low cut filter for removing a direct current component in the converted voice signal and a low frequency component unnecessary for voice recognition.

According to the above construction, the DC component in the voice signal passed to the voice recognition means and the low frequency component not necessary for voice recognition are removed, and the recognition rate is further enhanced.

The program recording medium of the second invention has recorded therein a voice data input conversion processing program for causing a computer to function as the voice input means, the input control means and the data conversion means in the first invention. Is characterized by.

According to the above structure, as in the case of the first invention, the frequency characteristic of the input voice is matched with the frequency characteristic of the learning data, and the recognition rate by the cutoff characteristic of the antialiasing filter included in the voice input means is obtained. Is prevented.

[0017]

BEST MODE FOR CARRYING OUT THE INVENTION The present invention will be described in detail below with reference to the embodiments shown in the drawings. FIG. 1 is a block diagram of the voice recognition device according to the present embodiment. This voice recognition device includes a voice input unit 11, an input control unit 12, and a data conversion unit 1.
3, the voice recognition unit 14 and the display unit 15 are roughly configured.

The voice input unit 11 converts the analog voice signal collected by the microphone 11a into an A / D converter 11b.
Digitize by. The audio signal digitized in this way is transmitted via the bus 16 to the data conversion unit 1
Input to 3. The input control unit 12 controls the audio input unit 11 to control the audio signal acquisition conditions such as the setting of the sampling frequency. Data converter 13
Is a digital voice signal captured by the voice input unit 11 by using a filter having the same characteristic as or a characteristic very similar to the cutoff characteristic of the antialiasing filter used when extracting the learning data for learning the acoustic model. Is subjected to a filtering process, and the filtered data is down-sampled to be converted into an audio signal having a predetermined frequency band.

The voice recognition unit 14 uses the voice recognition method using, for example, a hidden Markov model (hereinafter, referred to as HMM) to recognize the HMM corresponding to the recognition target word by using the voice signal converted by the data conversion unit 13. The occurrence probabilities are calculated for all of them, and the word corresponding to the HMM having the highest occurrence probability is used as the recognition result. Here, the HMM is a model of statistical characteristics of speech obtained from a large amount of speech data, and details of the speech recognition method using this HMM are described in "Speech Recognition by Probabilistic Model" by Seiichi Nakagawa. Have been described. The display unit 15 displays the recognition result in addition to the icon and the like.

FIG. 2 shows a flowchart of the voice data input conversion processing operation executed by the voice input unit 11, the input control unit 12, and the data conversion unit 13.
The audio data input conversion processing operation in the present embodiment will be specifically described with reference to FIG. Here, since a frequency band of 0 Hz to 6 kHz is used in the voice recognition process, a case will be described where sampling is performed at a sampling frequency of 12 kHz, which is twice the maximum frequency in the frequency band.

In step S1, the input control section 12 sets the sampling frequency in the audio input section 11 to 24 kHz, which is twice the target sampling frequency of 12 kHz, prior to voice acquisition. In step S2, the voice input unit 11 captures a voice signal at a sampling frequency of 24 kHz. In that case,
An audio signal that has passed through an anti-aliasing filter having a cutoff characteristic as shown by a broken line is obtained.

In step S3, learning data for learning an acoustic model (HMM) as shown by a solid line in FIG. 3 is applied to the voice signal taken in by the voice input unit 11 by the data conversion unit 13. The same characteristics as the blocking characteristics of the anti-aliasing filter used when extracting, or
A low pass filter with very similar characteristics is applied. Thus, the characteristics of the above transition region at the time of recognition are
The characteristics of the transition region at the time of learning the acoustic model (HMM) are made substantially the same.

In step S4, the data conversion section 13 further down-samples to 1/2 (that is, a voice signal having a target sampling frequency of 12 kHz used in the voice recognition process). In step S5, the data conversion unit 13 further applies a low-cut filter to remove DC components and low-frequency components (for example, components of 60 Hz or less) that are not necessary for voice recognition. After that, the voice signal is passed to the voice recognition unit 14, and the voice data input conversion processing operation ends.

The low-pass filter used in step S3 and the low-cut filter used in step S5 in the above flow chart are
For example, it can be easily realized by an IIR type filter or the like.

As described above, in the present embodiment,
The following voice data input conversion processing operation is added before the voice recognition processing by the voice recognition unit 14.
That is, the input control unit 12 sets the sampling frequency of the voice input unit 11 to 24 kHz, which is twice the sampling frequency 12 kHz used in the voice recognition processing by the voice recognition unit 14. Then, the data conversion unit 13
By applying a low-pass filter having the same characteristic as the cutoff characteristic of the anti-aliasing filter used when extracting the learning data or a very similar characteristic to the voice signal oversampled by the voice input unit 11, The characteristics of the transition band from the stop band to the pass band in the filter are set to be substantially the same as the characteristics of the transition band at the time of learning. Then, the data converter 13 further down-samples to 1/2 (that is, a voice signal with a sampling frequency of 12 kHz used in the voice recognition process).

Therefore, the frequency characteristic of the input voice and the frequency characteristic of the learning data used for learning the acoustic model can be matched by a simple preprocessing, and the recognition rate is lowered by the cutoff characteristic of the antialiasing filter in the A / D conversion circuit 11b. Can be prevented.

In the above embodiment, oversampling is performed at a sampling frequency twice as high as the sampling frequency used in the voice recognition processing.
Oversampling may be performed at a sampling frequency n times (n: an integer of 2 or more) (that is, 2n times the maximum frequency in the frequency band processed by the speech recognition process). However, in that case, it is necessary to perform 1 / n downsampling at the time of downsampling.

By the way, the functions as the voice input means, the input control means and the data conversion means in the above embodiment are realized by the voice data input conversion processing program recorded in the program recording medium. The program recording medium in the above embodiments is a program medium including a ROM (Read Only Memory).
Alternatively, it may be a program medium loaded in an external auxiliary storage device and read. In any case, the program reading means for reading the audio data input conversion processing program from the program medium may have a configuration of directly accessing and reading the program medium, or a RAM (random access memory). ) May be configured to be downloaded to a program storage area (not shown) provided in (1), and the program storage area may be accessed and read. The download program for downloading from the program medium to the program storage area of the RAM is assumed to be stored in the main body device in advance.

Here, the program medium is configured to be separable from the main body side, and is a tape system such as a magnetic tape or a cassette tape, a magnetic disk such as a floppy (registered trademark) disk or a hard disk, or a CD (compact disk)- Disk system of optical disk such as ROM, MO (magneto-optical) disk, MD (mini disk), DVD (digital video disk), card system such as IC (integrated circuit) card and optical card, mask ROM, EPROM (ultraviolet erasable type) RO
M), EEPROM (electrically erasable ROM), flash R
It is a medium that holds a program fixedly, including a semiconductor memory system such as OM.

Further, if the voice recognition device in the above-mentioned embodiment is equipped with a modem and is connectable to a communication network including the Internet, the program medium is fluid by downloading from the communication network or the like. Even if it is a medium that carries the program, it does not matter. In this case, the download program for downloading from the communication network is stored in the main body device in advance. Alternatively, it is assumed that the program is installed from another recording medium.

It should be noted that what is recorded on the recording medium is not limited to the program, and data can be recorded.

[0032]

As is apparent from the above, in the voice recognition apparatus of the first invention, the sampling frequency of the voice input means is controlled by the input control means so that the frequency band processed by the voice recognition means is in the pass band of the antialiasing filter. Set the oversampling frequency so that it is fully included,
Since the data conversion means performs the filtering process on the voice data sampled at the oversampling frequency by using the filter having the same characteristic or the similar characteristic as the cutoff characteristic of the antialiasing filter used at the time of extracting the learning data, In the pre-processing, the frequency characteristic of the input voice and the frequency characteristic of the learning data for learning the standard pattern or acoustic model can be matched. Therefore, it is possible to prevent the recognition rate from decreasing due to the cutoff characteristic of the anti-aliasing filter included in the voice input unit.

The voice recognition device of the first invention is
The input control means sets the sampling frequency to an oversampling frequency such that audio data having a frequency band n times (n: a positive integer of 2 or more) processed by the voice recognition means can be obtained. If you do
The 1 / n downsampling process by the data conversion means can be easily performed.

The voice recognition device of the first invention is
If the data conversion means is provided with a low-cut filter for removing a direct current component in the converted voice signal and a low-frequency component unnecessary for voice recognition, the direct current in the voice signal passed to the voice recognition means is reduced. It is possible to further improve the recognition rate by removing the components and low-frequency components unnecessary for voice recognition.

Further, the program recording medium of the second invention has recorded therein a voice data input conversion processing program for causing a computer to function as the voice input means, the input control means and the data conversion means in the first invention. As in the case of the first invention, the frequency characteristic of the input voice and the frequency characteristic of the learning data for learning the standard pattern or the acoustic model can be matched by a simple preprocessing. Therefore, it is possible to prevent the recognition rate from decreasing due to the cutoff characteristic of the anti-aliasing filter included in the voice input unit.

[Brief description of drawings]

FIG. 1 is a block diagram of a voice recognition device of the present invention.

FIG. 2 is a flowchart of a voice data input conversion processing operation executed by a voice input unit, an input control unit, and a data conversion unit in FIG.

3 is a diagram showing a filter cutoff characteristic at the time of recognition by the voice recognition device shown in FIG. 1 and a filter cutoff characteristic at the time of creating an acoustic model.

FIG. 4 is a block diagram showing an outline of a conventional voice recognition device.

FIG. 5 is a configuration diagram of an A / D conversion circuit.

6 is a diagram showing a filter cutoff characteristic at the time of recognition by the voice recognition device shown in FIG. 4 and a filter cutoff characteristic at the time of creating an acoustic model.

[Explanation of symbols]

11 ... voice input section, 11a ... Mike, 11b ... A / D converter, 12 ... Input control unit, 13 ... Data converter, 14 ... voice recognition unit, 15 ... Display, 16 ... bus.

─────────────────────────────────────────────────── ─── Continuation of front page (58) Fields surveyed (Int.Cl. ⁷ , DB name) G10L 21/00

Claims (4)

(57) [Claims] 1. A voice recognition device for recognizing an input voice by a voice recognition means, comprising an antialiasing filter having a pass band wider than a frequency band processed by the voice recognition means, and input voice having passed through the antialiasing filter. And a sampling frequency of the speech input means for sampling to obtain speech data, and the sampling frequency of the speech input means is set to an oversampling frequency such that the pass band of the antialiasing filter sufficiently includes the frequency band processed by the speech recognition means. The input control means to be set and a filter having the same characteristics as or similar to the cutoff characteristics of the antialiasing filter used at the time of extracting the learning data for learning the standard pattern or the acoustic model are sampled by the voice input means. Audio data In addition to the filtering process, the audio data after the filtering process is down-sampled and converted into audio data having a predetermined frequency band, and the data converting means is sent to the audio recognition means. Speech recognizer. 2. The voice recognition apparatus according to claim 1, wherein the input control unit has a sampling frequency n times as high as a frequency band processed by the voice recognition unit (n: a positive integer of 2 or more). A voice recognition device, characterized in that an oversampling frequency is set so that voice data having a frequency band can be obtained. 3. The voice recognition device according to claim 1 or 2, wherein the data conversion means removes a DC component in the converted voice signal and a low-frequency component unnecessary for voice recognition. A voice recognition device having a low-cut filter. 4. A computer-readable program recording medium on which a voice data input conversion processing program for causing a computer to function as the voice input means, the input control means, and the data converting means according to claim 1 is recorded.