SDR for audio source separation

[sophia1488]

Hi, I'm doing audio source separation (speech + pop music), and my target is speech, so there's no permutation problem.
I hope to evaluate with SDR, SIR, and SAR.
I saw the example in metrics.py, and I modified mix, clean, and est to be of the same shape.
I wonder if the result is reasonable. If so, I could use it directly, which is great!
Or should I modify mir_eval function?

from asteroid.metrics import get_metrics
mix = np.random.randn(2, 16000)        <--- 2: stereo (2-channel)
clean = np.random.randn(2, 16000)
est = np.random.randn(2, 16000)
metrics_dict = get_metrics(mix, clean, est, sample_rate=8000,
                           metrics_list='all')
pprint.pprint(metrics_dict)

Thank you!

[mpariente]

It'll compute the metrics for channels independently, right? I'm not sure this is what you want, is it?