E
STIMATIONNoises resulting from the engine, panel vibration, tire, wind, pass-by traffic, air conditioning, etc., could significantly degrade the music listening quality in a car.
This paper proposes a new system that makes use of a Noise Level Estimator (NLE) in tandem with s Dynamic Range Controller (DRC) to adaptively adjust the gain of an automotive audio system. A microphone is required in the NLE as the senor to pick
up music signals corrupted with cabin noise. The background noise level is estimated adaptively using the Least-Mean-Squares (LMS) algorithm. The a priori Signal-to-Noise Ratio (SNR) is calculated based on the noise level estimated above.
From the SNR, the gain of the audio input can be adjusted dynamically, with the aid of a static curve. The system has been implemented by using a Digital Signal Processor (DSP) on a real car. Results obtained from simulations and experiments reveal that the proposed system is capable of regulating the audio volume on the fly, in response to the noise in the car cabin.
4.1 Noise level estimation by least mean square method
System identification is the procedure of analyzing an unknown system. The well-known method to achieve the identification processing is the LMS algorithm [7].
The system structure is shown in Fig. 20, where x n( ) is the input signal and
p( )
x n is the desired signal. The desired signal can be determined by the optimal
coefficientw n , which is obtained by minimizing the error signal ( )s( ) e n . Basically, the system learns from its environment is designated as an adaptive filter where the filter coefficients are updated according to
( 1) ( ) ( ) ( ) the adaptation parameter and it should satisfy the condition
0 s 2 more time to converge to a minimum error, and vice versa.
Applied in a noisy condition, system identification is a pre-processing for NLE.
As the unknown plant is determined, the noise can be estimated from the difference
The FIR estimator for the system is defined by ( ) = ( ) s( )
y n x n w n . (24)
According to the processing of system ID the optimum parameters of the unknown plant can be determined by minimizing the Mean-Square-Error (MSE).
min [E eID2( )]n , (25) measured and the predicted system output,
eNLE( )n =d n( )−y n( )=x np( )+v n( )−x( )n ws( )n , (27)
The system setup is illustrated in Fig. 20. Here, ( )v n and x n are assumed to p( )
be uncorrelated. When the signal x n is determined and equal to the output p( ) ( )
y n then eNLE( )n is equal to ( )v n .
4.2 Dynamic range control
DRC of audio signal is used in many applications to match the dynamic behavior of the audio signal to different requirements. While recording, DRC protects the AD convector from overload or it is used in the signal path to optimally use the full amplitude range of a recording system. When reproducing music and speech in a car, the dynamics have to match the noise characteristic inside a car. A DRC is an automatic gain control device which modifies the dynamic range without introducing perceptible distortion. Fig. 21 shows a block diagram of DRC system.
After measuring the level of input signal x, the output signal y is affected by multiplying the delayed input signal by a factor g n( ) according to
( ) ( ) ( )
y n =g n x n⋅ −D , (28)
where D is the delay sample for non-process path and ( )g n is a gain factor that obtained according to the input level. Level measurement plays an important role in DRC. The rapidity of DRC depends also on the measurement of RMS values [10]. average coefficient. The transfer function is
( ) 1
attack or release status. If the input signal is lager than the previous signal, then system gives the attack coefficient AT. If the input signal is smaller than the previous signal, then system gives the release coefficient RT.
The relationship between input level and weighting level is defined by static curve.
In this paper, the output level and the weighting level are given as functions of the input SNR level g[dB] = f(SNR[dB]). Fig. 23 shows a static curve which varies with the SNR. According to the static curve, the limiter threshold is 6 dB. The output gain is limited when the SNR level exceeds the limiter threshold. All SNR levels less than this threshold lead to a constant output gain 6 dB. The SNR between 5~25 dB represents an amplifier, and has two slopes. The slope between 5 dB SNR to 20 dB SNR is 0.2. The slope between 20 dB SNR to 30 dB SNR is 0.3. Both the two parts are compressor curves. The SNR exceeding 30 dB leads to a constant gain 0 dB.
4.3 Integration of NLE and DRC modules
The two systems mentioned above have their own specific applications. This paper is focusing on how to improve the listening quality in a car environment. We combined NLE system and DRC system to deal with the noisy condition in a car.
The system block diagram is shown in Fig. 24. Take the NLE system as first step, the optimal parameter of the adaptive filter is obtained from eq. (19). Then according to eq. (25), eq. (26) and (27) the background noise level can be accurately estimated from a noisy signal. The signal then goes through RMS level measurement and turns into xRMS by eq. (29). Because the purpose is trying to adaptively adjust the output signal gain according to background noise, the signal to noise ratio (SNR) is an important factor. Use the power of background noise level and the power of signal level to calculate a priori SNR, where a priori SNR is given by
2 2
xp
ζ = v , (31)
where xp is the signal through plant and v is background noise.
As mentioned before, the static curve needs to be designed in advance. Based on a priori SNR and static curve, the output gain mapping can be determined.
4.4 System simulation and experimental investigation
Simulations and experiments are undertaken to validate the NLE and DRC modules proposed in the paper.
A. Simulations
We evaluated the proposed algorithm by performing a system simulation of the adaptive gain controller system. As the proposed processing mentioned before, the adaptive filter is used to track the plant when the NLE system is working.
According to the NLE system structure shown in Fig. 20, the filter coefficients are adaptively calculated from eq. (19), and the step size is equal to 0.45. From eq. (19), (25) and eq. (26) the optimal parameter is determined when the error is converging.
In Fig. 20, the program input x n( ) is an 8-second music excerpt and v n( ) is an 8-second background noise. The background noise is a level varying whitenoise.
When the unknown plant is determined, the background noise can be calculated by (27). See Fig. 25, the upper row is the original background noise and the lower row is the estimated background noise. After the background noise is estimated, the system then calculates the a priori SNR by (30). Fig. 26 shows the SNR curve, this is calculated by the signal and the background noise which is varying in 3 different levels. From Fig. 26, the SNR fluctuates between 12 dB to 31 dB. There is an obvious notch from the 2nd to the 6th second. According to the purposed algorithm, the output gain curve should obtain a higher gain when the SNR curve is falling to a
notch. The relationship between input SNR level and gain level is defined by a static curve. The static curve has described in section III. With the aid of the static curve and the calculated SNR curve, the output gain can be obtained. Fig. 27 shows the output gain curve. When the SNR is small, the system gets a higher gain. When the SNR is high, the system gets a lower gain.
B. Experiments platform of a fixed-point DSP, ADI BF-533, of Analog Device semi-conductor. The GRAS 40AC microphone with GRAS 26AC preamplifier was used for receiving the signal and the background noise. The position of the microphone is located at the center of the car.
The experiments were implemented by the following process. We turned on the system and microphone first. Through the DSP board, the program music was played form the right channel. With the ambient noise and the signal, the microphone picked up the noisy signal. The NLE system dealt with the noisy signal.
As soon as the adaptive filter approximated the plant, the background noise level can be estimated. After the noise level is estimated, the system would output gains for the input signal to maintain the SNR. There are two cases in this experiment. Case 1 used the ambient noise in a moving car. Case 2 used the level varying whitenoise.
Both the noises were played from the left channel. Fig. 29 shows the experimental result of case 1. Fig. 30 shows the result of case 2. The upper row is the waveform when the system is off. The lower row is the waveform when the system is on. In
case 1, the output signal changes moderate because the noise does not have clear level experimental results show that the system can dynamically adjust the audio gain when the background noise is varying.
R
EFERENCE[1]H. Drucker, “Speech Processing in a High Ambient Noise Environment,” IEEE Trans. Audio and Electroacoustics, vol. 16, no. 2, pp. 165-168, 1968.
[2] M. Heckmann, F. Berthommier and K. Kroschel, “Noise Adaptive Stream Weighting in Audio-Visual Speech Recognition,” EURASIP. Applied Signal Processing, vol. 11, pp. 1260-1273, 2002.
[3] D. E. Tsoukalas, J. N. Mourjopoulos and G. Kokkinakis, “Speech Enhancement Based on Audible Noise Suppression,” IEEE Trans. Speech and Audio Processing, vol.