【正文】 ivity Detector (VAD). The process of discriminating between the voice activity that is speech presence and silence that is speech absence is called voice activity detection. VAD algorithms typically extract some type of feature (. short time energy, zero crossing etc.) from the input signal and pared against threshold value, usually determined during speech absent period. Generally output of VAD algorithms is binary decision on a framebyframe basis having frame duration 2030 msec. A segment of speech is declared to contain voice activity (VAD = ‘1’) if measured value exceed a predetermined threshold otherwise it is declared a noise (VAD = ‘0’) figure 1shows VAD decisions. Several VAD algorithms were proposed based on various types of features extracted from the signal. Noise estimation can have major impact on the quality and Intelligibility of speech signal.Figure 1 shows VAD decisions [3]The early VAD Algorithms were based on energy levels and zero crossing [4], Ceptral features [4], the Itakura LPC spectral distance measures and the periodicity measures [2]. Some of VAD Algorithms are used in (GSM) System [3], cellular Networks [3], and digital cordless telephone systems [3]. VAD Algorithms are suitable for discontinues transmission in voice munication systems as they can be used to save the battery life of cellular phones. The majority of the VAD Algorithms encounter problems in low SNR conditions, particularly when the noise is nonstationary [1, 2]. Having an accurate VAD Algorithm in a nonstationary environment might not be sufficient in speech enhancement. Applications, as on accurate noise estimation is required at all times, even during speech activity. In case of Noise estimation algorithms they continuously track the noise spectrum therefore more suited for speech enhancement applications in nonstationary Scenarios.3. Classes of Noise Estimation AlgorithmsThere are three classes of noise estimation algorithms. Minimal tracking Algorithms, Time Recursive Algorithms and Histogram based Algorithms. All algorithms operate in the following fashion. First the signal is analyzed using short time spectra puted from short overlapping frames, typically 2030 msec. Windows with 50% overlap between adjacent frames. Then several consecutive frames called analysis segment are used in the putation of the noise spectrum. Typical time span of this segment may range from 400 msec. to 1 sec. The noise estimation algorithms are based on the assumptions that the analysis segment is too long enough to contain speech pauses and low energy signals segments and the noise present in the analysis segment is more stationary than speech, new assumption is that noise changes at a relatively slower rate than speech. The analysis segment has to be long enough to enpass speech pauses and low energy segments, but it also has to be short enough to track fast changes in the noise level, hence the chosen duration of the analysis segment will result from a trackoff between these two restrictions. Now we will see different classes of noise estimation Algorithms.Figure2 Plot of noisy speech power spectrum and local minimum [10] Minimal – Tracking AlgorithmsMinimal Tracking Algorithms are based on the assumption that the power of the noisy speech signal in individual frequency bands often decays to the power level of the noise, even during speech activity [12]. Hence by tracking the minimum of the noisy speech power in each frequency band, one can get a rough estimate of the noise level in that band. Two different algorithms were proposed for noise estimation first minimum statistics (MS) on noise estimation, which tracks the minimum of the noisy speech power spectrum within a finite window that is in analysis segment, and 2nd algorithm tracks the minimum continuously without requiring a window are explained in next section. Plot of noisy speech power spectrum and local minimum using (3) for a speech degraded by babble noise at 5dB SNR at frequency bin k=6 is shown in figure 2.. Minimum statistics (MS) Noise EstimationThe Minimum Statistics algorithm was originally proposed by Martin R. (1994) and later refined in [5] to include a bias pensation factor and better smoothing factor. Let y(n) = x(n) + d(n) denote the noise speech signal, where x(n) is the clean speech signal and d(n) is the noise signal, assume that x(n) and d(n) are statistically independent and zero mean. Noisy speech signal is transformed in the frequency domain by first applying a window w(n) to M samples of y(n) and then puting the Mpoint FFT of the windowed signal. (1)Where λ indicates the frame index and k the frequency bin index varient from k = 0, 1, 2 ... M1. Y(λ, k) is the short term Fourier Transform (STFT) of y(n). Periodogram of the noisy speech is approximately equal to the sum of periodogram of clean speech and noise given as| Y (λ, k) |178。 is the periodogram of noisy speed signal, |X (λ, k)|178。 fluctuates very rapidly over time, hence 1st under recursive version of periodogram can be used asP (λ, k) = α P (λ – 1, k) + (1α) |Y (λ, k) |178。( λ, k) = Bmin (λ, k). Pmin (λ, k)Figure3 Plot of true noise spectrum and estimated noise spectrum usingContinuous Spectral Minimum Tracking Arrows indicate regions wherenoise is overestimated [12] Continuous Spectral MinimumOne of the drawbacks of the minimal tracking employed in the MS algorithm is its inability to respond to fast changes of the noise spectrum [12]. A different method for tracking the spectral minima was proposed in [10]. In contrast to using a fixed window for tracking the minimum of noisy speech is in [5] the noise estimate is updated continuously by 。. We can obtain an estimate of the power spectrum of the noise by tracking the minimum of P(λ,k).Our finite window smoothing constant α chosen experimentally not too low or too high. There are two main issues with the spectral minimal – tracking approach the existence of