Recognise Elan media offset when reading annotation start/end times

[benfoley]

Elpis currently expects annotations to correlate to beginning of media. However, ELAN files can have offset media start times which would result in mis-aligned training.

When reading the annotation data from ELAN files, we should look in the header for media offset information and add/subtract that to the annotation times so that the alignment is correct.

Or trim the media sox input output trim 5000?

[benfoley]

@benfoley make an example ELAN file to demonstrate.

[benfoley]

The offset is the TIME_ORIGIN value.

        <MEDIA_DESCRIPTOR
            MEDIA_URL="file:///Users/bbb/Desktop/abui/abui-audio-1.wav"
            MIME_TYPE="audio/x-wav"
            RELATIVE_MEDIA_URL="./abui-audio-1.wav" TIME_ORIGIN="1000"/>

[benfoley]

Here's some sample data, adjusted the Abui toy corpus with various offset values.

abui-offset.zip