Enhancement test

[zeynepgulhanuslu]

Hi,

Thanks for sharing this code. I trained my enhancement model with my data. How can i clean audio files with my model. I couldn't find test code for applying enhancement. Could you help me?

Have a nice day,
Thank you

[nateanl]

Hi, after you get the trained model, you can create the input features for the audio and feed them to the model. Note that the input is a list of pytorch tensors and so is the output, like this:
[clean] = model([feature, noisy_mag])
clean is the clean magnitude, you can combine the noisy phase with the clean magnitude, then apply iSTFT method to get the clean audio signal. Thanks.

[max-krichenbauer]

[Edited after figuring it out]

Thank you for sharing the code!

I was stuck with the same problem (using edinburgh_tts_config).
I finally figured it out how to get the audio output:

output = self.model(input)
[embedding, mask_A, mask_B] = output
[one_hot_label, mag_mix, mag_s1, mag_s2, cos_s1, cos_s2] = label
mask_A = mask_A.cpu().detach().numpy()
mag_mix = mag_mix.cpu().detach().numpy()
clean_mag = np.multiply(mask_A, mag_mix)[0]
phase_mix = ### <- get the noisy phase
phase_mix = phase_mix[...,0] + 1j * phase_mix[...,1] # to complex array
sft_mat = np.multiply(clean_mag, phase_mix)
sig = librosa.core.istft(np.transpose(sft_mat), win_length=1024, hop_length=256)
librosa.output.write_wav("test_out.wav", sig, 16000)

[nateanl]

@max-krichenbauer Thanks for sharing the trick!
I plan to add the template for evaluating SDR for each configuration. I'll keep this issue open until it's done.
For your information, you can use the SI-SDR method in PyTorch.