Public evidence fixtures, prepared 2026-09-10 by Transkriba.

AMI natural meeting excerpt
Copyright: AMI Consortium. AMI Meeting Corpus EN2002a, Mix-Headset.
License: Creative Commons Attribution 4.0 International (CC BY 4.0)
https://creativecommons.org/licenses/by/4.0/
Corpus: https://groups.inf.ed.ac.uk/ami/corpus/
License and manual annotation release: https://groups.inf.ed.ac.uk/ami/download/
Audio source: https://groups.inf.ed.ac.uk/ami/AMICorpusMirror/amicorpus/EN2002a/audio/EN2002a.Mix-Headset.wav
Reference source: https://groups.inf.ed.ac.uk/ami/AMICorpusAnnotations/ami_public_manual_1.6.2.zip

Changes: cropped original audio [60,150) seconds to a 90-second WAV, retaining
16-kHz mono PCM. The excerpt was chosen before recognition. Only the beginning
of the large source WAV was downloaded, sufficient to include the entire excerpt.
ami-reference.json selects original manual word annotations starting within that
interval, shifts time by -60 seconds and caps word ends at the excerpt boundary.
It retains source speaker identifiers A/B/C/D and punctuation entries.
ami-meeting.json is the service backend's untouched segment output, with public
settings added and private job/storage/account identifiers omitted. TXT/DOCX are
exported from those segments without normalization or manual correction.

The English non-scenario meeting is a natural-speech example, not a Russian
accuracy benchmark, interview or customer testimonial. The English Whisper route
returned no speaker labels despite diarize=true; source speaker identifiers in
the reference are human annotations, not recognition output. Backend retrieval
required fixing its missing NumPy dependency; do not cite this as a latency test.
No endorsement by AMI Consortium is implied.

Subtitle import fixture
podcast-video.mp4 contains the existing synthetic Russian podcast demo and a
plain background. podcast-subtitled.mp4 imports podcast.srt with FFmpeg 8.0.1 as
a selectable mov_text track, copying H.264 video and AAC audio. These MP4 fixtures
were prepared outside Transkriba; the product exports SRT/VTT. The browser demo
loads the separate VTT track. They do not demonstrate burned-in captions or
compatibility with every editor. See the webpage for the exact import command.