Fig. 1. Flow diagram of proposed noise-estimation algorithm.

Fig. 2. Plot of noisy speech power spectrum and local minimum using (3) for a speech degraded by babble noise at 5 dB SNR at frequency bin k = 5.

Fig. 3. Top panel: Plot of estimated speech-presence probability based on the ratio Sr(λ, k). Bottom panel: spectrogram of the clean signal.

Fig. 4. Plot of true noise spectrum and the estimated noise spectrum using our proposed method for a speech degraded by babble noise at 5 dB SNR and single frequency f = 250 Hz.

Fig. 5. Comparison between the noise spectrum (for f = 1 kHz) estimated using the proposed algorithm (thick line) and Martin’s (Martin, 2001) (dashed line) algorithm for a sentence corrupted by car noise (t < 1.8 s) followed by a sentence corrupted by multi-talker babble (t > 1.8 s).

Fig. 6. Top panel: Plot of true noise spectrum and estimated noise spectrum using the proposed method for a noisy speech signal (5 dB SNR) at f = 250 Hz. Bottom panel: Plot of true noise spectrum and estimated noise spectrum using (Doblinger, 1995). Arrows indicate regions where noise is overestimated.

Fig. 7. Comparison of estimated noise spectrum (f = 500 Hz) of proposed method (dashed line) with that of Hirsch and Ehrlicher (1995) (solid line) for a noisy speech of SNR 20 dB (t < 1.8 s) followed by a noisy speech of SNR 5 dB (t > 1.8 s).

Fig. 8. Spectrograms of speech enhanced using Martin’s (2001) noise-estimation method (panel c), Cohen’s method (Cohen (2003)) (panel d) and the proposed noise-estimation method (panel e). Spectrograms of the clean and noisy speech signals are given in panels (a) and (b) respectively. Arrows in panels (c) and (d) at t > 3.8 s show the presence of residual noise due partly to the inability of the noise-estimation algorithms to track the sudden appearance of high-frequency noise in the last sentence (sentence 3). In contrast, as shown in panel (e), the residual noise is greatly reduced with the proposed noise-estimation algorithm.

Percent preference for the proposed method compared to other methods for single and mixed type noise

The normalized mean squared error (MSE) between the estimated and true noise spectra for various methods

Objective evaluation and comparison of the proposed noise-estimation algorithm in terms of segmental SNR values (dB) and LLR values