This article refers to the address: http://
Realistic voice communication can occur in noisy noisy environments. For example, cell phone communication in a factory can be affected by machine roaring; voice communication in the train cab can be disturbed by motor operation and rail collision. The statistical characteristics of noise are closely related to the scene. Even if the statistical characteristics of noise in the same situation may change with time, it is required that the noise canceling equipment must have the adaptive ability of noise tracking. After more than 40 years of development and improvement, the theory of adaptive signal processing has been applied in many fields [1, 2]. In this paper, the adaptive signal processing technology is applied to the noise cancellation of voice communication. The software and hardware parameters are optimized by simulation research and prototype experiment. An adaptive cancellation method suitable for mobile phone microphone and train cab is developed. system.1 noise cancellation principle
The principle of the adaptive noise cancellation system is shown in Figure 1. It has two inputs: the original input and the reference input, the reference input is the noise source ν1(n), and the original input is the noise-contaminated signal x(n)=s(n)+ν0(n). When the noise component ν0(n) is uncorrelated with the signal s(n) and related to the noise source ν1(n), the adaptive filter AF can adjust the coefficient of the self filter according to the error signal ej, so that the output yj tends to Î0(n) in the original input, so that error signal 0 tends to signal s(n).
The Least Mean Square (LMS) algorithm plays an important role in many adaptive signal processing algorithms because of its simple algorithm, small computational complexity and easy implementation [3, 4]. The system design of this paper uses the recursive method to implement the LMS algorithm.
The output of the adaptive filter AF at time j is expressed as:
Here, Wj is the filter coefficient at time j, and Xj is the filter input at time j. The weight coefficient of the next time (j+1) is adjusted according to the error signal ej of the current time j, and the recursive algorithm of the adjustment is:
Where μ is the step factor, which affects the convergence of the system. The sufficient conditions for system convergence are:
The μ value should be considered for the actual tradeoff of voice communication: when μ is too small, the weight coefficient converges slowly, and it cannot adapt to the occasion where the noise is not stable; when μ is too large, the cancellation effect is poor, and even the system diverges.
2 system simulation
2.1 Acquisition of sound samples
Matlab 7.0's Signal Processing Toolbox provides a Simulink module that reads audio data from a computer's standard audio device in real time, as shown in Figure 2.
The Signal To workspace in Figure 2 saves the acquired audio data to the Matlab workspace. As an example, this paper collected a human voice of about 0.3 s as the original signal, the sampling rate is 16 kHz, and its time domain waveform is shown in Figure 3.
2.2 Step factor optimization
The field noise is simulated with a sine wave of 250 Hz. Considering the noise situation at the scene, the value of the noise amplitude should be equivalent, which is 0.01. The sound samples and noise are superimposed as the original input of the system, as shown in Figure 4.
Taking μ=0.5, the 128-order adaptive filter is used for speech denoising. It can be seen from Fig. 5 that the output of the 0.03 s cancellation system is similar to the original signal.
The Wj of the adaptive filter needs to be iterated multiple times to reach the ideal value, that is, E[ej2]E trend to [Eej2)min requires a process. In view of the step size, it will significantly affect the cancellation effect of the system. The learning curve at different times is shown in the figure.
It can be seen from Figure 6:
(1) As the value of μ becomes larger, the learning speed of the system becomes significantly faster;
(2) When the signal is non-stationary, the excessive μ value is likely to cause the system to be out of adjustment, and the mean square error at 0.1 is significantly larger. Therefore, the value of μ should be compromised: on the one hand, when the noise is not stable, the learning time of the system should be less than the smoothing time of a tone (about 0.1 s); on the other hand, the system offset should be minimized. A fine comparison reveals that the learning time at μ=0.25 is already less than the smoothing time of one node, which is greatly improved when μ=0.1, so the long factor is 0.25 to meet the system requirements.
3 system implementation
3.1 Circuit design
The signal processor of the adaptive cancellation system uses TI's fixed-point DSP-TMS320VC5509. Its core clock can work stably at 200 MHz. It takes only one instruction cycle (5 ns) to complete two multiply-accumulate (MAC) operations.
The system has two microphone channels, one for the original input and the other for the reference input. The two-channel circuit form is exactly the same, and the signal conditioning and analog-to-digital conversion circuit is shown in Figure 7. The 16 b Codec AD73311 is connected to the DSP via a synchronous serial interface. The reference output of the AD73311 passes through the Buffer of the AD8058 as the DC bias of the microphone; the electret microphone signal uses an AC-coupled input with a gain of 50 to accommodate the dynamic range of the analog-to-digital converter. The AD73311 collects data periodically to trigger a receive interrupt, notifying the DSP to receive the data and process it accordingly.
3.2 software design
Adaptive filtering requires high real-time performance of the system, so the weight coefficient recursive, filtering, and noise cancellation are performed in the sampling interrupt service routine. The software design uses DSP instruction LMS dedicated to adaptive filtering, which can complete 2 parallel operations in one instruction cycle: multiply accumulate (MAC) and weight coefficient recursion. This instruction greatly improves code efficiency and enhances the real-time performance of adaptive signal processing.
The TI-DSP development environment CCS5000 provides support for the corresponding DSP library Dsplib. The library contains conventional digital signal processing functions including adaptive filtering, most of which are assembly code, and the code efficiency is high. Library functions provide C language function prototype declarations, allowing C programs to directly access, reducing the difficulty of digital signal processing programming.
The adaptive filtering library function strictly requires the alignment of the first address of the buffer. The software uses the pseudo-statement "#pragma DATA SECTION()" to specifically constrain the memory allocation of the cmd file, so as to make more reasonable use of the computing resources of the DSP chip. .
4 Conclusion
After repeated experiments, the prototype has been able to solve the denoising problem in voice communication under specific installation conditions.
In order to ensure that the noise components of the two channels have sufficient correlation, the installation pitch of the microphone should be less than 20 cm, and the speaker (signal) should be less than 5 cm from the original input. Otherwise, the system output will have a large amount of uncorrelated uncorrelated noise. Component.
It was found in the experiment that increasing the delay of the original input is beneficial to improve the cancellation effect of noise. After analysis, this is related to the working state of the adaptive filter. If the noise at the reference input is ahead of the original input signal (unsynchronized), the adaptive filter is actually a predictor, and the prediction is more difficult than filtering and not easy to accomplish. A simpler solution is to add software delay to the original input, which improves the ability of the cancellation system to adapt to the location of the noise source.
5W LED Underground Light,LED Light for Outdoor,LED Underground Light
Lead Semiconductor Technology Co., Ltd. , http://www.czdownlight.com