Noise Subspace Fuzzy C-means Clustering for Robust Speech Re(5)

Abstract. In this paper a fuzzy C-means (FCM) based approach for speech/non-speech discrimination is developed to build an effective voice activity detection (VAD) algorithm. The proposed VAD method is based on a soft-decision clustering approach built ove

Finally, the presence of a new cluster (speech frame detection) is satis ed if the following ratio holds:K 1F (l) = log 1/Kk=0 E(k, l) ¯ < mc >>γ(7)¯ where < mc > is the averaged noise prototype center and γ is the decision threshold. The set of noise prototypes are updated in pause frames (not satisfying equation 7)) using the adaptation of the standard FCM, replacing the oldest energy in the noise model, consisting of N samples, by the actual feature vector E(l). The initial prototype matrix M(l) at decision frame l is the previous one M(l 1), and the following update is applied to the fuzzy partition and prototype center matrices: uij(t+1)= 1/ =mi(t+1)until ||M (l) M (l)|| < f or i = 1, . . . , C, j = 1, . . . , NC 1/(1 m) l=1 (Dlj /Dij ) m (t+1) N Ej / j=1 uij (t+1) (t)N j=1uij(t+1)m(8)This sequential adaptation doesn’t involve high computational e ort although other kind of static adaptation rules could be applied. The algorithm described so far is presented as pseudo-code in the following: 1. Initialize Noise Model: – Select N feature vectors {E(i)}, i = 1, . . . , N . – Compute threshold γ. 2. Apply FCM clustering to feature vectors extracting C noise prototype centers {m(c)}, c = 1, . . . , C 3. for l=init to end (a) Compute E(l) over the MO window (b) if equation 7 holds then VAD=1 else Update noise prototype centers m(c) with equations 8. Figure 2 shows the operation of the proposed FCM-VAD on an utterance of the Spanish SpeechDat-Car (SDC) database [15]. The phonetic transcription is: “tres”, “nueve”,“zero”, “siete”, “µinko”, “dos”, “uno”, “otSo”, “seis”,“cuatro”. We also show the soft decision function and the selected threshold in the FCMVAD operation for the same phrase.4Experimental FrameworkSeveral experiments are commonly conducted to evaluate the performance of VAD algorithms. The analysis is normally focused on the determination of misclassi cation errors at di erent SNR levels [7], and the in uence of the VAD