Raw model outputs¶

What "raw" means¶

VTC's underlying model produces a frame-by-frame guess about who is speaking. The result is choppy: a single person talking can come out as several short detections with small gaps between them, which isn't very useful for analysis.

So before saving the results, VTC does two cleanups:

1. Drop very short isolated detections that are most likely model glitches.
2. Merge same-speaker detections that are separated by only a tiny gap (under 100 ms by default) into a single longer segment.

For most users, the cleaned-up output is what you want. But if you're debugging unexpected results, building your own custom post-processing, or studying the model's behaviour itself, the raw output is more informative.

How to keep the raw output¶

Add the --keep_raw flag when running VTC. The pre-cleanup detections will be saved alongside the cleaned-up output:

<output_folder>/
├── 📂 rttm/          # Cleaned-up segments (default output)
├── 📂 raw_rttm/      # Pre-cleanup, raw model detections (only if --keep_raw)
├── 📄 rttm.csv
└── 📄 raw_rttm.csv

The raw_rttm/ folder follows the same RTTM format as rttm/; it just contains more (and shorter) segments.

Technical Reference¶

Implementation details of the post-processing pipeline.

The default post-processing pipeline applied to model outputs:

1. Per-frame binarization of the four class scores against the active threshold set (f1.toml by default; hp.toml with --high_precision; or any TOML passed via --thresholds). See Threshold selection.
2. Drop short segments of duration less than --min_duration_on_s seconds (default 0.1).
3. Bridge same-class gaps of duration less than --min_duration_off_s seconds (default 0.1).

The segment-merging logic is the one from pyannote.audio.

Pass --keep_raw to additionally write the segments produced after step 1 but before steps 2–3 to <output>/raw_rttm/. To inspect even earlier, prior to binarization, use --save_logits (this can produce large files).