Closed jurihock closed 4 months ago
It looks like the DeepFilterNet
has been trained to use a fixed sample rate of 48 kHz, fixed window size of 960 samples (481 dft bins, which is not even a power of two) and fixed hop size of 480 samples (50% overlap). Changing any of this constraints requires retraining of the model. In addition, the author speaks of an algorithmic delay between 20 and 40 ms, which could also be a deal breaker.
Any chance to embed this into existing STFT procedure?