1 / 5

Training partially connected MLPs for MULTIBAND systems

Training partially connected MLPs for MULTIBAND systems. TCTS Faculté Polytechnique de Mons Belgium. Basic multiband system. Training independent neural networks for each subband. CBE subband 1. CBE subband 2. CBE subband 3. CBE subband 4. Posteriors subband 1. Posteriors subband 2.

kerem
Download Presentation

Training partially connected MLPs for MULTIBAND systems

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Training partially connected MLPsfor MULTIBAND systems TCTS Faculté Polytechnique de Mons Belgium RESPITE workshop - Sheffield

  2. Basic multiband system Training independent neural networks for each subband CBE subband 1 CBE subband 2 CBE subband 3 CBE subband 4 Posteriors subband 1 Posteriors subband 2 Posteriors subband 3 Posteriors subband 4 Subband posteriors or eventually outputs of the last hidden layer (NLDA) can then be combined to take a global decision. RESPITE workshop - Sheffield

  3. Partially connected MLPs If all the subband neural networks are trained on the same targets, can we train all the subband NN in one step ? CBE subband 1 CBE subband 2 CBE subband 3 CBE subband 4 NLDA Phoneme targets RESPITE workshop - Sheffield

  4. TEST 4 subband experiment on AURORA 2 data. The systems aretrained on white noise contaminated training data. System 1, fullband NN: 1 MLP: 300 x 1000 x 33 System 2, independent NN: 4 networks (2 hidden layers): 75 x 200 x 30 x 33 A MLP is used to recombine the outputs of the last hidden layer of each subband MLP. This MLP is trained on phoneme targets: 120 x 1000 x 33 System 3, partially connected NN: 1 partially connected network: 300 x 800 x 120 x 33 Input vector is the concatenation of the four subbands. A MLP is used to recombine the outputs of the last hidden layer. This MLP is trained on phoneme targets: 120 x 1000 x 33 RESPITE workshop - Sheffield

  5. TEST 4 subbands: 0-778 Hz, 707-1632 Hz, 1506-2709 Hz and 2122-4000Hz Test on the effect off added noise on the posteriors after subband recombination. Noise N1 (subway) at 0 dB in subband 2 only compared to clean speech with the 3 systems. Deviation of posteriors relative to clean speech (MSE) Word error rate – aurora test A, noise 1 RESPITE workshop - Sheffield

More Related