feat: Add audacity marker file support and creating annotation from rttm file

[yojul]

Problem

When evaluating speaker diarization pipelines, one might want to use Annotation objects and creating annotation from rttm (or other format) as well as serializing/writing annotation to rttm (or other format).
Audacity marker track feature is a very convenient (and free) way to create ground truth segmentation for speaker diarization. The format is a .txt file very similar to already implemented LAB file support but tab separated.

Solution

Refactor methods for various file format support

• Create a generic _serialize method to replace multiple to_{format} methods.
• Create a generic _write method method to replace multiple write_{format} methods.
• to_<format> and write_<format> are now partial methods from generic methods.

Currently supported formats are :

• Rttm : annotation.to_rttm() and annotation.write_rttm(file).
• Audacity : annotation.to_audacity() and annotation.write_audacity(file).
• Lab : annotation.to_lab() and annotation.write_lab(file).

Therefore, to add a new format one only need to implement _iter_{format} methods similarly to _iter_rttm or _iter_lab.

Creating annotation from audacity or rttm

Similarly to the from_df class methods, I created from_audacity and from_rttm class methods to create easily annotations from those file formats.

Usage :

with open('file.rttm') as f : 
      annotation = Annotation.from_rttm(f)

[hbredin]

Thanks for this PR.

Note that RTTM files may contain annotations for multiple audio files (hence the second uri field) in which case I am not sure what the Annotation.from_rttm method should do:

• raise an error?
• return a {uri: Annotation} mapping?

One okayish solution could be to add an option as_dict: bool = False to force returning a dict (second option) and raise an error if set to False and RTTM file contains multiple audio files...

[yojul]

Thank you for your feedback.

For more consistency with other "from" methods and the Annotation object itself, I suggest that from_rttm works as follow :

• if no uri is specified, the default_uri is taken from the first line of the rttm file. Then, if there is more than one uri in the file, it raises an Exception asking to specify a uri
• if uri is specified as parameter, from_rttm is only reading the lines with the specified uri.

Thus, it insures consistency with Annotation uri and rttm uri and that from_rttm is still creating a single Annotation object (as other similar methods).

I also added a condition to only read lines starting with "SPEAKER" corresponding to speech segments.