feeding information on annotations from a csv (non-rttm like) format

[alecristia]

Description

We're working on calling pyannote.metrics (specifically, to get identification error rate) from within R (package tutorial), so we are hoping to use as few python packages & other adaptations as possible. Our annotations are in a csv, which contains the following info for each annotated segment:

• onset
• offset
• talker

Is it truly necessary to create rttm-like files? Or could we hack in as in the code stub below?

Steps/Code to Reproduce

This code doesn't run, it is just a stub of what the code would look like:

from pyannote.core import Annotation, Segment
reference = Annotation()

# read csv

# create a segment matrix, with things like (sorry, bash syntax cameo!):
# reference[Segment($onset, $offset)] = $talker

from pyannote.metrics.diarization import Identification
metrics["ider"] = identification.IdentificationErrorRate(parallel=True)

# write csv with results

Thanks in advance!

[hbredin]

Yes. No need to create an RTTM file.

from pyannote.core import Segment, Annotation

manual_reference = Annotation()
for onset, offset, speaker in ...:
    manual_reference[Segment(onset, offset)] = speaker
automatic_hypothesis = Annotation()
for onset, offset, speaker in ...:
    automatic_hypothesis[Segment(onset, offset)] = speaker

from pyannote.metrics.identification import IdentificationErrorRate
metric = IdentificationErrorRate()
error_rate = metric(manual_reference, automatic_hypothesis)
print(error_rate)