KR19980037008A - Remote speech input and its processing method using microphone array - Google Patents

Abstract

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a remote speech input apparatus using a microphone array and a remote speech input processing method therefor in a field of application of a computer-based speech recognition system, In order to solve the problem that many inconveniences have arisen due to careful attention, the present invention has a multi-channel microphone array, receives voice signals having different time delay values from the multi-channel microphones simultaneously, An automatic speech detector for detecting a speech signal from multi-channel signals input from a microphone array; a time delay estimator for estimating time delay information of the detected inter-channel speech signals; And a time delay and a signal for adding a voice signal of all channels after offsetting a time delay existing in a voice signal between channels to generate a voice having an improved signal-to-noise ratio using the estimated time delay information;

The microphone and speaker have been developed to have a function of automatically receiving a voice having a signal-to-noise ratio of a level (25 dB) or more at a distance of about 40 to 80 cm, even when the speaker does not pay attention to the position of the microphone, Is used as the input of speech recognition system.

Description

A remote audio input device using a microphone array and its remote audio input processing method

1 is a block diagram of a remote audio input apparatus according to the present invention;

2 is a flowchart of the entire remote speech input processing of the present invention;

3 is a detailed flowchart of time delay estimation according to FIG.

4 is a detailed time-delay-signal addition flowchart according to Fig.

Description of the Related Art [0002]

10: Microphone array part 11: Microphone array

12: A / D signal converting section 20: automatic voice detecting section

30: Time delay estimator 40: Time delay -

Most speech recognition systems currently being developed and announced adopt a method of inputting a voice using a single microphone.

However, in the case of receiving a single micro voice, careful attention must be paid to the position of the microphone at the time of voice input. Particularly, in the case of a speech recognition system in which a high signal-to-noise ratio must be maintained, a microphone must be placed close to the speaker. Therefore, it is inconvenient for many people to use it practically.

The remote audio input device of the present invention is designed to solve such a problem.

That is, it is developed to have a function of automatically receiving a voice having a signal-to-noise ratio equal to or higher than a permissible level (25 dB) when a microphone and a speaker are spaced by about 40 to 80 cm and do not pay attention to the position of the microphone Respectively. In order to satisfy this, a noise elimination technique using a microphone array was applied.

[Object of the invention]

It is an object of the present invention to enable a speech recognition system to have high speech recognition performance by allowing a user to maintain a high signal-to-noise ratio of an input speech signal without paying a great deal of attention to the position of a microphone.

[TECHNICAL FIELD OF THE INVENTION AND RELATED ART OF THE SAME]

The present invention relates to speech recognition using a computer and, more particularly, to a microphone array having a plurality of micro-arrays arranged in front of a speaker, receiving a speaker's voice simultaneously through a plurality of microphones, and performing multi-channel beamforming ) Technology to reproduce a noise-reduced voice, thereby enabling a voice recognition system to input a clear voice whose noise is reduced even when it is uttered at a relatively long distance from the microphone, and a remote voice input device using the microphone array And a voice input processing method.

Conventional technologies in this field include a voice input technique using one channel and a multi-channel microphone voice input technique.

Currently, most speech recognition systems that have been developed and announced adopt a method of receiving a voice using a single microphone. However, in the case of receiving a single micro voice, the user must pay close attention to the position of the microphone at the time of the ignition. In particular, in the case of a speech recognition system that maintains a high signal-to-noise ratio, the microphone must be placed close to the speaker. Therefore, there are various inconveniences that ordinary people use in practice.

Further, when reproducing voice signals input from a plurality of channels using a conventional multi-channel beamforming technique as voice signals with reduced noise, it is difficult to accurately detect the time delay between voice signals of various channels .

[Technical Problem]

Accordingly, in order to solve the above-mentioned problems, the present invention has been made to solve the above-mentioned problems, and it is an object of the present invention to provide a speech recognition system that inputs a clear voice to a voice recognition system by maintaining a high signal- A center clipping method is applied to a speech signal to increase cross-correlation between speech signals of adjacent channels and a cross-correlation is used between speech signals of adjacent channels. In addition, Prior to this, we added a noise level normalization procedure to develop the best signal-to-noise ratio in a given situation.

[Structure and operation of the invention]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, the present invention will be described in detail with reference to the accompanying drawings.

FIG. 1 is a block diagram of a remote audio input apparatus according to the present invention. The microphone array unit 10 includes multi-channel microphones. And an adder 40.

As shown in FIG. 1, the microphone array unit 10 includes a microphone array 11 composed of a plurality of (or multi-channel) microphones, a microphone array 11 receiving a voice signal having different time delay values from microphones located at various places, And an analog-to-digital converter (12).

The automatic speech detection unit 20 detects a speech part necessary for estimating time delay information between speech signals input from microphones of each channel, from an input signal.

The time delay estimating unit 30 estimates time delay information between the audio signals of each channel detected by the automatic sound detecting unit 20.

The time delay-signal adder 40 performs a function of canceling a time delay existing in a voice signal between channels, and then adding voice signals of all channels to generate a voice having improved signal-to-noise ratio.

2 is a flowchart of the overall remote speech input processing of the present invention.

This is accomplished by converting a speech signal input from a micro-array of micro-channels of a plurality of channels into a digital signal (S10), detecting only the speech signal portion of each channel thus converted (S20) (S30). Then, the speech signals having different time delays on a channel-by-channel basis are processed to have the same time delay, and speech signals of all channels are added to output speech (S40).

The above-described time delay estimation process (S30) will be described in more detail with reference to FIG.

The time delay information is detected by extracting a speech signal of a predetermined interval at each channel and then estimating the time delay information between the speech signal of the reference channel and the speech signal of the other channel by obtaining a cross-correlation function.

In step 31 (S31) of extracting a speech frame, a speech frame is extracted from an automatic speech detection apparatus and a speech frame composed of N speech samples is extracted for each channel simultaneously.

Step 32 (S32) is a process of detecting a maximum value and a minimum value, and obtains a maximum value and a minimum value of the speech samples for each of the speech frames composed of N speech samples extracted for each channel.

Step 33 (S33) is a center clipping process. In order to increase the accuracy in the estimation of the time delay information, a center clipping function is applied to each voice frame for each channel. Center clipping takes from 0 to 50 to 80% of the maximum and minimum of speech frames for speech samples. That is, the expression is as follows.

[Equation 1]

y (n) = 0, for C min ≤x (n) ≤C max, n = 0, ..., N-1

y (n) -x (n), for elsewhere, n = 0, ..., N-1

Here, x (n) is a voice sample values, y (n) is a clipping level center clipping taking the function output value, C _min and C _max are respectively fixed ratio to the minimum and maximum values of the speech samples belonging to a speech frame (50 To 80%).

In step 34 (S34), in the process of calculating the cross-correlation function, a reference channel is first determined for each center-clipped voice frame, and a crosscorrelation function between the voice frame of the reference channel and the voice frame of the other channel is determined. Respectively. The formula of the cross-correlation function between the voice of the reference channel and the voice of the other channel is as follows.

&Quot; (2) "

Where x ₀ represents the speech samples of the reference channel and x _k represents the speech samples of the other channel. Also, M represents the number of channels, and N represents the size of a voice frame.

Step 35 (S35) is a time delay estimation process. After obtaining the maximum value of the obtained cross-correlation functions, the time τ is determined as a time delay value between the reference channel and the other channel. This can be expressed as follows.

&Quot; (3) "

Here, k denotes other specific channels except for the reference channel, and t3 denotes a cross-correlation function between the voice signal of the reference channel and the voice signal of the other channel k.

The above-described time delay-signal addition process (S40) will now be described with reference to FIG.

In order to cancel the time delay existing between the reference channel and the other channel, the voice signals of all the channels are synchronized in time by delaying by a negative value of the time delay value, And serves to generate the final improved voice signal.

Step 41 (S41) is a process of canceling the time delay existing on a channel-by-channel basis, and delays the voice signal by a negative value of the time delay value so that there is no time delay of the inter-channel voice signals.

Step 42 (S42) is a process of normalizing the difference of the noise level for each channel, and multiplying the voice signal by the noise level normalization coefficient for each channel. The channel-specific noise level normalization coefficient W _k is obtained as follows.

&Quot; (4) "

Here, k denotes a specific channel of the microphone array, and t5 denotes a noise signal in a non-speech interval between the channel k and the reference channel, respectively. Also, N is the size of the non-speech noise frame.

Step 43 (S43) is a process of canceling the time delay for each channel and adding normalized voice signals over all channels and dividing by the number of channels. The items of the above-mentioned steps 41 to 43 can be expressed by the following equations.

&Quot; (5) "

Where M denotes the number of microphone array channels, and τ _k denotes the time delay between the reference channel and the other channel k.

The operation of the present invention will now be described.

In the present invention, instead of using one microphone, several microphones are used for inputting voice regardless of the position of the microphone. If you distribute multiple microphones on the front of the speaker, you can expand the input range. In order to generate a clean speech signal of reduced noise from the voice signals of various channels inputted from the microphone array 11 composed of a plurality of microphones, voice signals having different time delays for each channel are processed to have the same time delay Next, a method of adding voice signals of all channels is used. If the time delay of the audio signal is processed to be the same across all channels and added, the audio signals of the respective channels having the same waveform are superimposed on each other, so that the size of the waveform is proportional to the number of channels, The power component of the speech signal, which is defined as the sum of squares of the number of channels, increases in proportion to the square of the number of channels. On the other hand, when the noise components existing together with the speech signal in each channel have the characteristics of statistical independence in time, the components of noise are not completely overlapped even if the components of each channel are added because of the inherent randomization characteristic of noise. Therefore, the power component in this case increases in proportion to the number of channels. Therefore, the signal-to-noise ratio of the speech signal (the power component of the speech signal / the power component of the noise) is equal to the number of channels. Therefore, if the noise between each channel is statistically independent, finally the signal-processed voice signal gains gain of the signal-to-noise ratio as many as the number of microphones.

In the present invention, the signal-to-noise ratio of a voice signal is improved by using such a method.

When this method is used, it is necessary to accurately detect the difference in time delay of the audio signals between the respective channels.

In order to solve this problem, in the present invention, first, an automatic speech detection unit 20 detects a speech signal portion among signals input to respective channels. In order to measure the time delay between the voice signals for each channel, the time delay estimator 30 estimates the interchannel time delay using the voice frames per channel detected by the automatic voice detection unit 20.

The maximum / minimum value detection process, which is a technique applied by the time delay estimator 30, is used to set the center clipping level required in the next center clipping process. In the center clipping process, the portion inside the center clipping level is assigned as 0 in the waveform of the audio signal, and only the outside portion is left as the original waveform value. By doing so, it is possible to obtain a much more accurate result when the inter-channel time delay is estimated by using the next process, i.e., the cross-correlation, by removing a waveform having a small size, which is relatively noise-

In the calculation process of the cross-correlation function, a cross-correlation function is calculated for two different voice frames. The cross-correlation function calculates the function value while moving the voice frame from the two channels back and forth on the time axis. The position of the overlapping of the two voice frames (the voice frame of one channel having no time delay shifted by the time delay of the other channel voice signal ) Represents the maximum value. In the next step, the time delay estimation process, a position having a maximum value is found in the obtained cross-correlation function, and this point is estimated as a time delay value.

After the time delay estimation process is completed, the time delay-signal is subjected to inverse delay processing to cancel the time delay for each channel by the divider 40, and after the normalization process of the voice signal of each channel, The signal is added to the completely superimposed waveform on the object. In this case, the normalization process normalizes the signal level of the channel so that the noise level of all the channels is the same, and the voice signal generated in this case can obtain a gain of the highest signal-to-noise ratio.

[Effects of the Invention]

While the present invention has been described with reference to the preferred embodiments thereof, it will be apparent to those skilled in the art that various modifications and variations can be made in the present invention without departing from the spirit or scope of the invention. It is possible to input a voice of a good quality with less noise and to improve the performance of the voice recognition system.

Claims (9)

A microphone array unit 10 having a multichannel microphone array for simultaneously receiving voice signals having different time delay values from the microphones of the multichannel and converting the digital signals; An automatic speech detection unit 20 for detecting a speech signal from multi-channel signals input from the microphone array unit 10, A time delay estimator (30) for estimating time delay information of the detected interchannel audio signals; And A time delay and a signal for adding voice signals of all channels after offsetting a time delay existing in a voice signal between channels to generate a voice having an improved signal-to-noise ratio using the estimated time delay information, And the remote audio input device using the microphone array. The method of claim 1, wherein the number of channels of the microphone array is And is configured to be proportional to a power component of a voice signal to a power component of a noise. 2. The apparatus of claim 1, wherein the time delay estimator (30) Channel time delay is estimated using a channel-by-channel speech frame. A remote speech input processing method having a microphone array, A first step of converting a voice signal input from a micro-array of micro-channels of a plurality of channels into a digital signal; A second step of detecting only the speech signal part of each of the converted channels; A third step of estimating a time delay between the detected voice signals for each channel; And And a fourth step of processing the voice signals having different time delays for each channel to have the same time delay by the estimation and adding voice signals of all channels to output voice. 5. The method of claim 4, wherein the third step A first step of extracting a speech frame composed of N speech samples simultaneously for each channel; A second step of obtaining a maximum value and a minimum value of each of the extracted voice frames for each channel such as a voice sample; A third step of performing center clipping on the voice frame for each channel for accurate estimation of the time delay information; A fourth step of determining a reference channel for the center-clipped voice frame for each channel, and then obtaining a cross-correlation function between the voice frame of the reference channel and the voice frame of the other channel; And Calculating a maximum value of the obtained cross-correlation functions, and then estimating the time as a time delay value between the reference channel and the other channel. 6. The method of claim 5, wherein the center clipping in the third step Wherein the speech signal has a size ranging from zero to a predetermined ratio of a maximum value and a minimum value of a voice frame for voice samples. 5. The method of claim 4, wherein the fourth step A first step of canceling a time delay existing for each channel; A second step of normalizing a noise level difference for each channel after the cancellation; And A third step of adding the normalized voice signals over all channels and dividing the normalized voice signals by the number of channels by canceling the time delay for each channel. 8. The method of claim 7, wherein the first step And delaying the voice signal by a negative value of the time delay value so that there is no time delay of the inter-channel voice signals. 8. The method of claim 7, wherein the second step comprises: And normalizing the voice signal by multiplying the voice signal by a noise level normalization coefficient for each channel.