Diarization from 4 (default softformer) speakers to 2)

[QuentinFuxa]

"""
PSEUDO ALGORITHM to Transform Diarization that predicts up to 4 speakers, to constrain it to predict up to 2 speakers

this can be done by tricking the output:
Ouput in self.total_preds is of shape x, x, 4. We want x, x, 2

So each iteration i, we have a ranking of the 4 speaker prediction, from 1 to 4

If 2 speakers, the model will have closed predictions for the 2 positions of speaker 1, same for speaker 2.
we should not just do active_speakers = np.argmax(preds_np, axis=1), but take top 2, and build the mapping 4:2 . It may evoluate, as second detected speaker can still be speaker 1, or be speaker 2.
Use np.argsort(preds_np, axis=1)[:, -2:] to get the top 2.

DETECT SPEAKER TOP 1 FOR PREDICTION i : DS_a_{i}
DETECT SPEAKER TOP 2 FOR PREDICTION i : DS_b_{i}

ATTRIBUTED SPEAKER FOR PREDICTION i : AS_{i}

GROUND TRUTH SPEAKER A : GTS_A
GROUND TRUTH SPEAKER B : GTS_B
DIST(a, b) : distance between detected speakers a and b.

i
0: AS_0 <- A
1: AS_1 <- B
2:
    IF DIST(DS_a_0, DS_a_1) < DIST(DS_a_0, DS_a_2) AND DIST(DS_a_0, DS_a_1) < DIST(DS_a_1, DS_a_2):
        #likely that DS_a_0 = DS_a_1
        AS_1 <- A
        AS_2 <- B
    ELIF DIST(DS_a_0, DS_a_2) < DIST(DS_a_0, DS_a_1) AND DIST(DS_a_0, DS_a_2) < DIST(DS_a_1, DS_a_2):
        AS_2 <- A
    ELSE
        AS_2 <- B

3: ...
"""

[QuentinFuxa]

This remains unimplemented. I opened #410 to track a tested max_speakers contract. The work needs stable cross-chunk speaker identities and real two-speaker evaluation before exposing the option. Closing this discussion in favor of the issue.