TWI396186B - Speech enhancement technique based on blind source separation for far-field noisy speech recognition - Google Patents

Description

Remote Noise Speech Recognition Based on Blind Signal Separation Speech Enhancement Technology

The invention relates to a long-distance noise speech recognition system established by combining an independent component analysis method and a subspace speech enhancement method, and proposes a related art application in the field of speech processing, such as noise cancellation, speech enhancement, and speech recognition technology. The architecture of the noise recognition system combined with noise filtering and speech enhancement and recognition processing in noise speech.

In terms of speech separation identification, the present invention proposes Blind Source Separation (BSS) speech enhancement technology to improve the long-distance speech recognition rate, and separately separates the speech signal and the noise signal in the mixed signal to have In the noisy environment, there is also a good recognition effect.

For the current common speech recognition technology and the patent technology related to the Republic of China in recent years, in 2007, Shen Jialin et al. invented the system and method for using sound recognition to select sound content, in addition to this patent. In addition to the text content and voice information stored in a source database, the identification module of the patented technology utilizes an acoustic model for speech recognition. The acoustic model comparison method is a Hidden Markov Model (HMM), Neural Networks and Dynamic Time Warping (DTW) for speech recognition, but the identification module of this technology is only suitable for close-range speech recognition, and it also has a significant impact in noisy environments. The identification module identification rate of this patent.

In addition, in 2009, the voice-activated voice controller of the invention, which was invented by Chen Maolin, includes a wireless transmitter and a wireless receiver, and the wireless transmitter and receiver are RF transmitters and RFs. The receiver is connected to the voice recognition component, and the wireless receiver is connected to the control component. In the identification technology, the voice message is pre-emphasized, the window processing is processed, the linear prediction coefficient is processed, and finally, the signal is processed. The spectral coefficient process processes the voice message through the speech recognition component for comparison with whether the voice message is In accordance with the voice command, in this invention, it is only suitable for voice recognition in a quiet environment, and is not directed to processing in a noisy environment. In the same year, the "voice recognition device, system and method" invented by Zhang Jianyang et al., in the electronic data processing device of the speech recognition system, uses several different hierarchical databases to store different voice data. The invention includes a receiving mode. The group, the conversion module, the cutting module and the identification module mainly convert the received voice signal into a digital signal, and then cut the digital signal into a plurality of voice data having a chronological order according to a cutting rule, and then The matching speech data is retrieved in the database, and the technique is also free for speech processing in a noisy environment.

Looking at the speech recognition systems mentioned above, they are mostly close-range speech recognizers, that is, the distance between the speaker and the microphone should not be too far, because when the distance between the two is farther away, the influence of noise and long-distance will be This makes the recognition effect worse, and it is also less convenient because of the limitation of the need to use it at close range. In addition, for the environment containing noise, the speech recognition system of the above invention patent is also significantly affected, so how to remove unnecessary noise signals in the real environment is also an important technical content of the present invention.

The present invention focuses on setting up a voice control system for the Ubiquitous environment, using two different noise environments to set up two systems, such as the second and fourth systems, and the second figure uses a microphone mixer. To construct a Ubiquitous environment operation, while a large range of microphones are arranged for recording, if a multi-channel recording card is used, there is a high cost. The present invention proposes to use a microphone mixer to collect sound, and to collect and mix signals of multiple microphones into one. A single signal, which reduces the amount of computation (only a single signal can be processed), and can reduce the cost of construction. The subspace enhancement method is used to deal with the problem of noise, and finally the result is sent to the HTK (Hidden Markov Model Toolkit) speech recognizer.

In addition, an environment for a large interference source is proposed. For example, the architecture of the fourth figure uses two-stage speech enhancement to remove the interference source signal, and uses the blind signal separation and subspace enhancement method to remove the large interference source and improve the speech recognition. The recognition rate of the device.

For the characteristics and characteristics of the technology, the system uses a microphone array composed of two microphones to perform sound collection, which reduces a large amount of construction cost and signal processing operation complexity, and can effectively improve the degree of noise interference at a long distance, so as to enhance The subsequent voice signal is suitable for speech recognition. Therefore, the present invention can be applied to digital home voice control devices to make family life more digital and convenient.

For the comparison of the general speech recognition system, the system uses a microphone array composed of two microphones to collect the sound, which can effectively reduce the construction cost, reduce the calculation amount, and perform voice enhancement for the noise interference of the long-distance use environment. Therefore, compared with the general short-distance speech recognition system, not only the convenience is improved, but also the identification system is more robust, which can effectively improve the convenience and correctness of the current traditional speech recognition.

In addition, the papers in the appendix: "Remote Noise Recognition Based on Blind Signal Separation Speech Enhancement Technology" is attached to the present invention. The attachment was published in the "Republic of China Computing Language" on September 1-2, 2009. The website of the Society is available at http://www.aclclp.org.tw/rocling/rocling2009_c.php and the disclosure is incorporated by reference in its entirety.

The system proposes a blind signal separation speech enhancement technology to deal with interference caused by a long-distance noise environment. The technical flow chart of the long-distance noise speech recognition system of the present invention is mainly shown in the first figure, and can be mainly divided into two stages: voice enhancement. Stage, speech recognition stage.

In the voice enhancement stage: firstly, the audio is collected through the dual-channel microphone array, and the collected two-channel signal is transmitted to the PC through the Line-In interface and sampled by the sound card, and the two-channel signal is separated by the blind signal. Initially separate noise and speech signals. As shown in the second figure.

Here, we use FastICA as the algorithm for blind signal separation, and the pre-processing part is divided into two steps: centralized variable and whitening. The processing steps of centralized variable are mainly to deduct the average value of the mixed signal. Solve the solution after simplification The solution process of the mixed matrix is calculated as follows, where x represents the mixed signal.

The second step of preprocessing is data whitening. The purpose of data whitening is to make the converted data non-correlated and the Variance value is one. If the conversion data is z, The Covariance matrix of this data becomes the identity matrix. Therefore, the whitening of the data is to find a whitening matrix V, and linearly convert the received signal x and make the covariation matrix into a unit matrix.

z = Vx , E { zz ^T }= I (2)

The FastICA core algorithm can be divided into objective function and optimization algorithm. The objective function we choose the negative entropy (Negentropy) function, and the optimization algorithm uses the Newton method to quickly obtain the convergence value. The definition of entropy (Entropy) can be expressed by the following formula according to the discrete signal or continuous signal: H ( y )=-Σ P ( y )log P ( y ) (3)

H ( y )=-ʃ f ( y )log f ( y ) dy (4)

In the voice signal part, when the signal y is Gaussian, its entropy is the maximum value. Therefore, for the convenience of calculation, we use negative entropy as the basis. As shown in formula (5), y _gauss has the same variation matrix as y. The Gaussian distribution signal, so when the signal y is Gaussian, the negative entropy is zero. To simplify the calculation, we simplify the formula (5) to the formula (6).

J ( y )= H ( y _gauss )- H ( y ) (5)

Where G is the control equation, the signal v is the Gaussian distribution signal with the mean value of zero and the variance is one, here we set E{G(y)}=E{G(W ^T x)}, W is the de-mixing matrix, x is a mixed signal, so equation (6) can be rewritten into formula (7). When E{G(W ^T x)} is maximum, the non-Gaussian distribution of the highest voice signal can be found, and finally the Newton method is used. On behalf of the operation, the solution matrix M is solved.

W ← E { xG ( W ^T x )}- E { G '( W ^T x )} W (8)

The voice signal separated by FastICA still carries residual noise. Therefore, we use the Signal Subspace Speech Enhancement Method to process the residual noise, as shown in the third figure.

Assuming that the original mixed signal is the original voice signal subspace y plus another noise signal subspace n _S , we must find a filter F so that the mixed signal can be filtered to obtain a clean signal y'=Fz. The filtered signal is compared with the original signal to calculate the error of the filter. The error value δ is calculated as follows:

Where δ _y represents the distortion of the speech signal filtered by the filter. Indicates the distortion caused by the noise that is not filtered. To optimize the filter in the signal subspace, the degree of speech distortion is minimal for the voice signal portion, and the residual noise for the noise signal portion. As long as it is as close as possible to the extent that it does not affect the identification result, instead of requiring no residual noise components, the noise signal is eliminated by analyzing the speech signal and background noise by Eigen-Decomposition.

The voice signal after sub-space speech enhancement needs to be processed by endpoint detection (VAD). Endpoint detection is a very important step before speech recognition, because endpoint detection can remove redundant invalid segments, otherwise the invalid segment will seriously affect the speech recognition rate. The sub-energy and zero-crossing rate are used to capture the effective segment of the speech signal, and the mute segment is removed so that the speech signal is suitable for speech recognition. As shown in the fourth figure.

Where x(n) represents the amplitude energy of the nth sample point, and x(n-1) represents the previous sample point, so the zero-crossing rate means that there are different sign times between two consecutive samples. After the correct voice signal is removed, the identification can begin.

Speech recognition phase: Finally, in the speech recognition part, we use HTK to train the speech model, speech feature capture and speech recognizer, as shown in the fifth figure. The experimental results of the long-distance noise speech recognition system of the present invention are shown in the sixth, seventh and eighth figures.

Figure 1: Technical flow chart of the long-distance noise speech recognition system of the present invention

Figure 2: Flow chart of the rapid independent component analysis method of the present invention

Figure 3: Schematic diagram of the subspace speech enhancement technology of the present invention

Figure 4: Flow chart of endpoint detection processing technology of the present invention

Figure 5: Flow chart of the speech recognition technology of the present invention

Figure 6: Application interface diagram of the long-distance noise speech recognition system of the present invention

Figure 7: Signal-to-Noise Ratio and Segmented Signal-to-Noise Ratio (SNR) performance of the long-range noise speech recognition system of the present invention in various background noise environments

Figure 8: The recognition rate of the long-distance noise recognition system of the present invention for various background noise environments

Appendix: Submission of the present invention: Long-range noise speech recognition based on blind signal separation speech enhancement technology

Claims (1)

A long-distance noise speech recognition system technology based on blind signal separation speech enhancement technology, the flow of implementation steps is as follows: Before using the independent component analysis method to obtain the de-mixing matrix, it is necessary to assume that the signal sources are independent of each other, however, In reality, the signal sources are not independent of each other. Therefore, before performing the independent component analysis process, the pre-processing must be performed before the de-mixing matrix can be found. The pre-processing method of the system of the present invention is Centering. And data whitening (Whitening) processing: (1) centralized variable calculation And (2) data whitening calculation z = Vx , E { zz ^T } = I ; according to the central limit theorem, the multiple non-Gaussian and independent signals are added together, which will make the whole tend to Gaussian distribution. Therefore, if any two random signals tend to be non-Gaussian, the components of the two signals are independent of each other. In estimating the non-Gaussian signal part, the present invention uses Negentropy to evaluate the calculation, and the best. The algorithm uses Newton's method to quickly obtain convergence values: (3) Negative entropy calculation And (4) Newton's method of iterative calculation W ← E { xG ( W ^T x )}- E { G '( W ^T x )} W ; in the long-distance noise speech recognition system of the present invention, the mixed signal is the original The voice signal subspace plus another noise signal subspace, since the present invention uses an estimate in the time domain, the voice signal distortion filtered by the filter is calculated directly in the time domain. And distortion caused by noise that is not filtered out: (5) Calculation of filter error value The enhanced voice signal is then processed by the endpoint detection method to capture the effective segment of the voice signal through the energy and zero-crossing rate, and the mute segment is removed to make the voice signal suitable for speech recognition: (6) Point detection calculation And in the speech recognition part using the HTK voice kit for identification.