search

breizhn / DTLN

Tensorflow 2.x implementation of the DTLN real time speech denoising model. With TF-lite, ONNX and real-time audio processing support.
MIT License
567 stars 160 forks source link

How to retraining for 10ms audio #20

Closed DyncEric closed 3 years ago

DyncEric commented 3 years ago

How to retraining this model for 10ms audio frame : ）

breizhn commented 3 years ago

Hi, you have to change self.blockLen and self.block_shift inside the DTLN_model constructor to your desired values or call the properties inside your training script before building the model. For 10ms block length set blockLen to 160 (samples). As block shift 5ms should work, so set it to 80 (samples). The quality is probably similar to the 32 ms model.

DyncEric commented 3 years ago

Thanks， i will try

cloudvc commented 2 years ago

hello there,

I modified the block Len to 160, block shift to 80 and running the training, but the loss is nan...