Demonstration
of Time-domain Speech Dereverberation based on
Maximum Likelihood Estimation with Time-varying Gaussian Source Model
Reference
- Nakatani, T., Juang, B.H., Yoshioka, T., Kinoshita, K., Delcroix, M., and Miyoshi, M., ``Speech dereverberation based on maximum likelihood estimation with time-varying Gaussian source model,'' submitted to IEEE Trans. Audio, Speech, and Language processing.
Task
- Dereverberation of female/male utterances
- Dereverberation filter was estimated from each reverberant sound shown below, and each dereverberated sound was obtained by applying the dereverberation filter to the reverberant sound.
- Length of the dereverberation filter in each channel: 3000 taps (fixed)
- Sampling frequency: 8 kHz
Source models used for performance
comparison
|
WG |
Stationary white Gaussian model |
|
TVWG |
Time-varying white Gaussian model |
|
TVAR |
Time-varyign autoregressive Gaussian model |
|
TVARC |
Time-varying autoregressive Gaussian model with codebook prior |
|
TVGC |
Time-varying Gaussian source model with codebook prior |
Demonstration sounds
- Dereverberation of reverberant female speech with reverberation time (RT60) of 0.5 sec.
|
Source
model |
Source
sound |
Reverberant
sound |
Dereverberated sound |
|
WG |
|
|
|
|
TVWG |
|
||
|
TVAR |
|
||
|
TVARC |
|
||
|
TVGC |
|
- Dereverberation of reverberant female speech using TVGC under different reverberation time conditions.
The length of the dereverberation filter was set at 3000 taps, which is much larger than that required for the strict inverse filtering under the 0.1 sec reverberation time condition, and much smaller than that required for the strict inverse filtering under the 1.0 sec reverberation time condition.
|
Rtime (second) |
Source sound |
Reverberant sound |
Dereverberated sound |
|
0.1 |
|
|
|
|
0.5 |
|
|
|
|
1.0 |
|
|
- Dereverberation of noisy reverberant female speech using TVGC.
The reverberation time (RT60) was 0.5 sec, and the reverberant signal to noise power ratio was 30 dB. The noise was stationary white Gaussian noise. A post-processing was also performed for eliminating musical noise that remained in the dereverberated signal.
|
Rtime (second) |
Noisy reverberant sound |
Sound denoised using Wiener filter |
Denoised and Dereverberated sound |
Post- processed |
|
0.5 |
|
|
|
|
4. Dereverberation of male speech captured by a stereo voice recorder in a noisy reverberant conference room.
Note that no noise reduction was performed for this example to show the robustness of our dereverberation method, and thus a certain level of noise remain audible after the dereverberation.
|
RT60/SNR |
Source sound |
Reverberant sound |
Dereverberated sound |
|
Unknown |
N/A |
|
|