1. ABSTRACT: Noise cancellation systems have been implemented to counter the effects of echoes in communications systems. These systems use algorithms that have been implemented using digital signal processors to track how a noise signal is changed by the frequency response of the communication system it travels in and then use that information to cancel that noise. These systems have been used for vocal communication systems, but the idea can be expanded to the use of music. In a recording studio, a singer usually stands in a sound booth and sings with headphones playing the instrumental that the vocals are accompanying. This is a non-ideal situation as the singer cannot hear her own voice along with the instrumentals. The goal of this project is to allow for the recording of a voice with instrumentals in the background and to isolate just the singer’s voice. This noise cancellation system will take a known instrumental signal, output it over a loudspeaker, while sending it to a digital signal processor (DSP) that will convolve it with the estimated frequency response of the room. A microphone will pick up the singer’s voice, along with the instrumental signal in the background that has traveled throughout the room and will send it to the DSP. The DSP will subtract the known instrumental signal convolved with the estimation of the room’s frequency response from the recorded microphone signal. The anticipated result of the system is the singer’s voice plus the background instrumental attenuated by 20 dB. Noise Cancellation System University of Pennsylvania Department of Electrical & Systems Engineering GROUP 14: Anil Makhijani, EE ’05 & Jeevan Puthiamadathil, EE & WH ’06 ADVISORS: Prof. Dan Lee & Yuanqing Lin DEMO TIMES: Thursday, April 21st, 2005 9:30 – 11:30AM SYSTEM DIAGRAMS & OVERVIEW: DOUBLE TALK DETECTOR Without Singing (No Double Talk Detected) With Singing (Double Talk Detected) c(n)Instrumental that is saved in the “Audio File” s(n)Singer’s voice H(n)Frequency response of the room G(n)Adapting filter’s current estimate of the frequency response of the room E(n)= s(n) + c(n) * H(n) – c(n) * G(n) = s(n) + c(n) * [H(n) – G(n)] M(n)= s(n) + c(n) * H(n) Notes:1) Function of t means the signal is analog; Function of n, signal is digital. 2) The “ * “ is the convolution operator. Without Singing: An instrumental (programmed as an audio file on the DSP) is sent to a D/A converter which sends the signal to a speaker which plays the instrumental. Simultaneously, the signal is sent to the nLMS Adaptive Filter. The microphone picks up the sound in the room (the instrumental convolved with the room’s frequency response), passes it through an A/D converter, which then sends the signal to the Double Talk Detector and the linear subtractor. The Double Talk detector then sends out a “no double talk” signal to the filter and subtractor. During this time, the filter keeps adapting to get its estimation of the frequency response as accurate as possible (an error signal sent from the subtractor back to the adaptive filter is minimized). With the no double talk signal, the linear subtractor stops subtracting and a zero signal is sent to the file. With Singing: The instrumental (programmed as an audio file on the DSP) is sent to a D/A converter which sends the signal to a speaker which plays the instrumental. Simultaneously, the signal is sent to the nLMS Adaptive Filter. The microphone picks up the sound in the room (the voice of the singer and the instrumental convolved with the room’s frequency response), passes it through an A/D converter, which then sends the signal to the Double Talk Detector and the linear subtractor. The Double Talk detector then sends out a “double talk” signal to the filter and subtractor. During this time, the filter stops adapting as the error signal (otherwise minimized) contains the voice of the singer. As voice is detected, the subtractor resumes subtracting and a signal containing the voice and the attenuated instrumental is sent to the file. Detector should detect when the singer is and is not singing Ideally, the detector outputs “true” when S(t) is zero and “false” when the output is non-zero Double Talk Detector Algorithm The most common algorithm used is the Geigal Algorithm The Geigal Algorithm defines variable ζ as: ζ(n) = |M(n)| / (max(|C(n)|, |C(n-1)|, |C(n-2)|, … |C(n-N)|) User-picked threshold value T ζ > T - there is double talk ζ < T - there is no double talk T is figured through experimentation ADAPTING FILTER ALGORITHM Normalized Least Mean Squared (nLMS) G(n+1) = G(n) + (μ*C(n)*E(n)) / (C(n)*C T (n)), where C T (n) is defined as the transpose of C(n) μ: amount the filter will change after each adaptation High μ: filter changes by a big amount each iteration – can result in overshoot Small μ: slow convergence of filter Past sound cancellation  0.3<μ<0.5