`separation.bss_eval_sources()` and `separation.bss_eval_sources_framewise()`

[Yossarian0916]

Although the docs describe the differences between these two functions, it's still unclear, even the returned results from these two functions have different shapes. Could someone elaborate the differences and the usage of separation.bss_eval_sources() and separation.bss_eval_sources_framewise()?

[craffel]

@faroit Any thoughts?

[faroit]

bss_eval_sources_framewise computes several bss_eval_sources as set by the window and hop parameter. if your signals are short: use non-framewise, if your signals are long >10s, use the framewise metric.

See the main loop in the code to understand whats going on:

mir_eval/mir_eval/separation.py

Lines 591 to 601 in 19ed4df

	for k in range(nwin):
	win_slice = slice(k * hop, k * hop + window)
	ref_slice = reference_sources[:, win_slice, :]
	est_slice = estimated_sources[:, win_slice, :]
	# check for a silent frame
	if (not _any_source_silent(ref_slice) and
	not _any_source_silent(est_slice)):
	sdr[:, k], isr[:, k], sir[:, k], sar[:, k], perm[:, k] = \
	bss_eval_images(
	ref_slice, est_slice, compute_permutation
	)