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
585 stars 161 forks source link

Some questions about real-time denoising #37

Open yolin329 opened 3 years ago

yolin329 commented 3 years ago

Breizhn, I have few question of real-time denoising.

First, what the 'quantization' argument in 'convert_weights_to_tf_lite.py' mean. If i want ti get the tf lite model, I just use the tf model (.h5) I just trained as input to 'convert_weights_to_tf_lite.py', right? And, if I set the argument 'quantization' be True, then the result of denoising shold be better, right? Did I misunderstanding?

And, when I run the code, 'real_time_dtln_audio.py', I notice that result of output is much worse than the result when I use 'run_evaluation.py' with tf model. Is that because I used tf lite model without quantization?

The last but not least, if I want to change the value of block_len_ms and block_shift_ms, I need to retrain the model with new value of batchsize, blockLen, and block_shift in DTLN_model(), right?

Thank you a lot.