JPS6219760B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech recognition device, and is particularly characterized in that it prevents erroneous recognition due to external noise.

Extract the feature parameters of the audio signal, use these feature parameters to recognize the voice of a certain speaker,
2. Description of the Related Art Various speech recognition devices have been proposed in the past that output signals for controlling external devices in accordance with recognized speech. An example of this is shown in the block diagram of FIG. an input section 1 including an acoustic-electrical signal converter such as a microphone that converts input audio into an electrical signal;
An input band filter 2 that extracts only the frequency band to be processed from an input acoustic signal, a feature extractor 3 that extracts features of the audio signal, and a standard pattern that stores pre-registered audio features as a standard pattern. The main components are a storage unit 4, a recognition processing unit 5 that identifies the input voice by comparing the characteristic pattern extracted from the input voice with a standard pattern, and an output control unit 6 that outputs the recognition processing result. An input signal amplitude normalization circuit 7 for improving the recognition rate, a time axis adjustment section 8, and a registration control section 9 that is in charge of processing when registering standard patterns of voice features in advance are added to the system.

There are many methods that can be considered as parameters for extracting voice features, such as frequency spectrum distribution, correlation function, number of zero crossings, formant frequency, or linear prediction coefficient. The so-called filter bank method, which separates and extracts patterns and checks the correlation with a standard pattern, is often used as a method that can obtain a high recognition rate with a relatively simple configuration.

Now, in such a voice recognition device, how do you determine the S/N of the control command voice that the device should recognize?
An extremely important issue is how to input signals with a good ratio, that is, with a high signal to noise ratio. When considering the environment in which speech recognition devices are actually used, it is rather rare that they operate in a quiet place without noise, and usually speech recognition devices are exposed to various ambient sounds other than control command voices. It is used in situations where there is a lot of noise from the perspective of the device. In other words, the ability to separate and extract control command voices from various sounds is a major factor in determining the recognition performance of the device.

Most of the frequency components of human speech are contained within the audio band of 100Hz to 5KHz, and conventional speech recognition devices perform recognition processing using the spectral distribution within this band as a feature parameter. By the way, the band of the frequency spectrum distribution exhibited by the various ambient sounds mentioned above is much wider than the band of the voice, and includes both components within the voice band and higher frequency components. is the majority. In the conventional speech recognition device having the configuration shown in Fig. 1, the input band filter 2 takes in the signal in the voice band, and there is no distinction between the control command voice and the components of the ambient sound within the voice band. . Therefore, the signal inputted from the input band filter 2 to the subsequent feature extraction section 3 is a signal in which the control command voice signal and the voice band component are added, and the feature extraction section 3 extracts features from the signal in which both are added. Extract parameters. However, the voice band components of the ambient sounds are noise to the voice recognition device whose level, frequency components, and duration change every moment, so the control command voice and the irregularly mixed ambient sounds are added. If registration and recognition operations are performed using parameters extracted from the combined signals, recognition performance will deteriorate. As long as the control command voice is used as a signal and the ambient sound is used as noise, the S/N ratio is sufficiently good, there is no problem. Conventional equipment also uses various methods to ensure this good S/N ratio. Measures have been taken. For example, it is a simple and reliable method to use a close-talking microphone as the input section microphone. A close-talking microphone, also called a noise-cancelling microphone, is constructed so that it can be used close to the lips to suppress the infiltration of ambient sounds. The adoption of this close-talking microphone improves the S/N ratio to a certain degree, and is highly effective in practice. However, it tends to be less effective against acoustic signals that have large amplitudes and include relatively high frequency components among ambient sounds.
However, it cannot be said that it is sufficient.

The present invention has been made for the purpose of preventing deterioration of the recognition performance of such a speech recognition device due to ambient sounds.

In other words, it attempts to determine whether or not a "sound" should be recognized or registered based on the difference in frequency spectrum distribution between a person's voice and surrounding noise.
It focuses on frequency components above Hz and examines the duration of these frequency components to clarify the differences between control command voices and ambient sounds.

Hereinafter, an embodiment of the present invention will be described in detail with reference to FIG. 2. 1 has an input section and is composed of a microphone 10 and an amplifier 11. After the signal from this input section 1 passes through the input band filter 2,
Enter feature extraction section 3. The feature extraction unit 3 includes N bandpass filters (hereinafter abbreviated as BPF) 30, 31,...3 whose center frequencies are ₁ , ₂ ... _N , respectively.
2 an integrator 33 that integrates the outputs of each of these filters;
34,...35 A multiplexer 36 that switches the output of each integrator, and an analog-digital (A/D) converter 37 that converts the output level of each filter that has passed through the multiplexer into a digital signal.
Consisting of: In addition, N BPF30,3
It is assumed that the band handled by all of the input band filters 1, . The A/D converter 37 converts each filter component of the signal from the input section 1 at appropriate time intervals (in most cases, 10 msec).
After being sampled (before and after) and converted into a digital code, it is stored in the storage memory 5 via a microcomputer 50 including an I/O port of the recognition processing unit 5.
1 is stored. Another standard pattern memory 4 is connected to the microcomputer 50, and the characteristic patterns of control command voices are stored in advance in association with the control contents. In the voice recognition mode, the control voice is input as described above,
The signal strings extracted by the filters 30, 31, . The input speech is specified by calculating the patterns and determining the standard pattern with the smallest difference. In general, human speech is not always the same over time even when the same language is uttered, so it is well known that some kind of time axis adjustment circuit 8 must be added as shown in Figure 1. However, since this time axis adjustment can be processed by software simultaneously when the microcomputer 50 calculates the characteristic pattern, it is omitted in the embodiment shown in FIG. Further, in the present invention shown in FIG. 2, the amplitude normalization circuit 7 in FIG.
is also omitted for the same reason.

Speech is always captured in the recognition mode, and when the input speech is interrupted, that is, during a pause period, the above-mentioned recognition calculation is executed, and the previous input speech is identified by the pattern matching method. At this time, when the input voice can be specified, that is, when the input voice matches some standard pattern within an allowable error range, the microcomputer 50 controls the output control unit 6 to respond to the control voice. Outputs signals for controlling external devices. If the input voice cannot be specified, the external device control signal is not output, and the display unit 12 is driven, for example, to notify that the input voice could not be specified, and prompt the speaker to speak again.

Next, in the configuration shown in FIG. 2, how the discrimination between the audio signal and the surrounding noise, which is the content of the present invention, is performed will be explained.

The signal from the input section 1 enters the input band filter 2 and is also input to the high pass filter 14 at the same time. This high-pass filter 14 passes components of the acoustic signal from the input section 1 that have a band equal to or higher than that of the input band filter 2 . However, although it is called a high-pass filter, it is not necessary to pass frequencies higher than the highest frequency coming from the microphone 10, so it has high-frequency attenuation characteristics that match the characteristics of the microphone 10. The signal component passing through this high-pass filter 14 determines the difference between the control command voice and the ambient sound. After the signal from the high-pass filter 14 is integrated by a high-frequency signal integrator 15, it is input to a duration detector 16 to check the duration of its occurrence.

As mentioned above, most of the components of the audio signal are concentrated within the band limited by the input band filter 2 and the BPFs 30, 31, ...32, and the components within the band controlled by the high-pass filter 14 are Very few. On the contrary, various noises occurring in the surroundings contain many components within the band of this high-pass filter 14. Therefore, it is thought that it is possible to determine whether a sound is a sound or an ambient sound based on the presence or absence of components within this band. However, there are consonants such as |S|, |K|, and |Z| that contain components in this band, and it may be dangerous to simply make a judgment based on the presence or absence of components, so to speak, and the level. Recognize. Therefore, in the present invention, the duration detection circuit 16 is provided to incorporate elements in the time axis direction in order to make accurate determinations. Normally people do the above |S|, |K
When consonants such as | and |Z| are uttered, the duration of the in-band components of the high-pass filter 14 contained in these consonants is within a range of about 20 msec to 150 msec. Therefore, the duration detection circuit 16 detects the duration of the signal component passing through the high-pass filter 14, and the duration detects the duration of the signal component passing through the high-pass filter 14.
If the noise does not appear within 150 msec, it is determined that this is noise other than voice, and a signal indicating this is given to the microcomputer 50. microcomputer 5
Upon receiving this signal, 0 stops taking in audio data and waits for the next audio input.

As described above, according to the present invention, the characteristic of the difference in duration of band components in the discrimination between speech and ambient sound is used, and the discrimination is high, which greatly contributes to improving the recognition rate of speech recognition devices. It is. Moreover, its configuration is very simple and its practicality is extremely high.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing a conventional speech recognition device, and FIG. 2 is a block diagram showing the configuration of the speech recognition device of the present invention, in which 1 is an input section, 3 is a feature extraction section, and 4 is a standard pattern storage section. , 5 is a recognition processing unit,
Reference numeral 14 indicates a high-pass filter, and 16 indicates a duration detection circuit.

Claims (1)

[Claims] 1. an acoustic-electrical signal conversion means, an audio band filter that passes audio band components of the electrical signal from the conversion means, and a high-pass filter that passes high frequency components above the audio band of the electrical signal; and a duration detection means for detecting the duration of the signal component passing through the high-pass filter, and the duration detection means detects whether the signal passing through the high-pass filter has a duration other than the duration of normal speech. A voice recognition device characterized by interrupting voice recognition operation if the operation lasts for a long time.