Demos for paper published on IEEE/ACM Transactions on Audio, Speech, and Language Processing: A Regression Approach to Speech Enhancement Based on Deep Neural Networks [PDF]

*Yong Xu, Jun Du, Li-Rong Dai and Chin-Hui LEE, Fellow, IEEE

E-mail: xuyong62@mail.ustc.edu.cn, *Yong Xu

 

Section 1: More enhanced speech demos based on Deep neural networks

Section 2: Related waveforms refered in the submitted paper

 

Section 1: More enhanced speech demos based on Deep neural networks

Selected results on some unseen noise types at the common SNRs (10dB, 5dB):

	Improved DNN	LogMMSE	Noisy	Clean
Machine gun, SNR10	wav	wav	wav	wav
Machine gun, SNR5	wav	wav	wav	wav
Buccaneer1, SNR10	wav	wav	wav	wav
Buccaneer1, SNR5	wav	wav	wav	wav
Destroyer engine, SNR10	wav	wav	wav	wav
Destroyer engine, SNR5	wav	wav	wav	wav
Factory1, SNR10	wav	wav	wav	wav
Factory1, SNR5	wav	wav	wav	wav

 

Testing on the unseen noise type: Pink at different SNRs below, and the clean speech (female, TIMIT) is the same.

	Improved DNN	LogMMSE	Noisy
SNR20	wav	wav	wav
SNR15	wav	wav	wav
SNR10	wav	wav	wav
SNR5	wav	wav	wav
SNR0	wav	wav	wav
SNR-5	wav	wav	wav

 

Testing on the unseen noise type: Exhibition at different SNRs below, and the clean speech (male, TIMIT) is the same.

	Improved DNN	LogMMSE	Noisy
SNR20	wav	wav	wav
SNR15	wav	wav	wav
SNR10	wav	wav	wav
SNR5	wav	wav	wav
SNR0	wav	wav	wav
SNR-5	wav	wav	wav

 

 

Section 2: Related waveforms refered in the submitted paper

(a) DNN-baseline

(b) clean

(c) noisy

Fig. 4 in the submitted paper to indicate the over-smoothing problem in the DNN-based method: Spectrograms of an utterance tested with AWGN at 0dB SNR: for the (a) DNN-baseline estimated (left), (b) the clean (middle) and (c) the noisy (right) speech.

 

(a) noisy speech

(b) LogMMSE

(c) DNN baseline

(d) improved by dropout

(e) improved by GV

(f) improved by dropout and GV

(g) jointly improved by dropout, NAT and GV equalization

clean speech

Fig. 8 in the submitted paper to compare the performance of different proposed systems and the traditional LogMMSE method: Spectrograms of an utterance tested with Exhibition (unseen) noise at SNR= 5dB. (a) noisy speech (PESQ=1.42), (b) LogMMSE (PESQ=1.83), (c) DNN baseline (PESQ=1.87), (d) improved by dropout (PESQ=2.06), (e) improved by GV equalization (PESQ=2.00), (f) improved by dropout and GV (PESQ=2.13), (g) jointly improved by dropout, NAT and GV equalization (PESQ=2.25), and the clean speech (PESQ=4.5).

 

(a) 104-noise DNN enhanced

(b) LogMMSE enhanced

(c) 4-noise DNN enhanced

(d) noisy speech

Fig. 9 in the submitted paper to indicate the ability of suppression against non-stationary noise : Spectrograms of an utterance tested on Machine gun (unseen) noise at SNR= -5dB: with (a) 104-noise DNN enhanced (PESQ=2.78), (b) LogMMSE enhanced (PESQ=1.86), (c) 4-noise DNN enhanced (PESQ=2.14), and (c) noisy speech (PESQ=1.85).

 

(a) DNN enhanced

(b) LogMMSE enhanced

(c) noisy

(d) clean

Fig. 10 in the submitted paper to indicate the ability of suppression against severely changing noise envrionments (non-stationary) : Spectrograms of an utterance corrupted in succession by different noise types tested on changing noise environments at SNR = 5dB: (a) DNN enhanced (PESQ=2.99), (b) the LogMMSE enhanced (PESQ=1.46), (c) noisy (PESQ=2.05), and (d) clean speech (PESQ=4.50).

 

(a) improved DNN

(b) LogMMSE

(c) DNN model further post-processed using the LogMMSE method

(d) noisy

Fig. 11 in the submitted paper to indicate the ability of enhancing the real-world noisy utterance : Spectrograms of a noisy utterance extracted from the movie Forrest Gump with: (a) improved DNN, (b) LogMMSE, (c) DNN model further post-processed using the LogMMSE method, and (d) noisy speech