Segmentation

[wmcmurtry]

Hello,

There is misalignment between the documentation within the run-segmented.sh script and the userguide page. I wanted to write this post in case anyone else was having trouble using their own .seg file to do transcriptions. I am using kaldi and eesen-offline-transcriber on their own, as opposed to coupling them together with Vagrant. I am using aspire models also.

Format:
The run-segmented script says to use a ".stm" or a ".cha" file while the userguide specifies how to format a .seg file. You may be able to use either of the first two, but I created a .seg file, following the instructions in the userguide.

Script:
The run-segmented.sh script includes the line cat build/trans/$basename/$basename.stm | grep -v "inter_segment_gap" | grep -v "ignore_time_segment_in_scoring" | awk '{OFMT = "%.0f"; print $1,$2,$4*100,($5-$4)*100,"M S U",$2}' > build/diarization/$basename/show.seg
I think this is intended to convert a stm file to a seg format, but if you already have your seg file formatted correctly, you can replace this line with cp build/trans/$basename/$basename.stm build/diarization/$basename/show.seg
Lastly, if you are using aspire, make sure to change the next line to make --file=Makefile.aspire SEGMENTS=show.seg build/output/$basename.{txt,trs,ctm,sbv,srt,labels}

File locations:
Have your .wav file in src-audio and your .stm file (which is in seg format as described in userguide, not in stm format. We just have the .stm at the end because that is what the script is looking for.) in the same directory from which you are calling the script. There may be other ways to do this, but this is what works for me.

Running the script:
do include the file extension
./run-segmented.sh myfile.wav

I used pyAudioAnalysis to segment my files.

Overall, a different segmentation method had negligible impact on my transcription output. It took a lot of time to do and I would not recommend it. The LIUM segmenter seems to work fine.