basic format discussion (rttm)

[alecristia]

Pro's rttm

• standard
• well defined
• validated over a very long time
• richest format
• we could use a single evaluation for all diarization tasks, including SAD

Con's rttm in divime

• poorly implemented right now (but this should be fixed anyway)
• we could use the convention but for the transcript part, we'd have all time stamps with asterisk
• not easy to read by target users (but this could be fixed by generating other formats)
• unnecessarily complete for several tasks (speech detection)

We discuss alternative formats:

• stm NIST (for transcriptions) file, channel, speaker, beg, (dur/end), category [male, far... properties], transcription
• ctm NIST (for phones)
• WCE: own format
• NOTE including more formats means more complexity

CONCLUSION:

• [ ] fix our use of rttm and make it standard for sad/vad, talker, and role diarization, and VCM --> all using same eval scripts
• [ ] for WCE, as well as input for these, we'll use stm http://www1.icsi.berkeley.edu/Speech/docs/sctk-1.2/infmts.htm#stm_fmt_name_0
• we are still waiting to see how we eval WCE

Note: Check Coconut for conversion across formats

[alecristia]

after further discussions, we decide VCM will also write its output to the "speaker ID" column, and thus use eval from evalDiar