Raw model outputs¶

What "raw" means¶

The model makes a decision for every short frame of audio, so its output is fragmented: one person speaking can appear as several short segments separated by small gaps. Before writing the results, VTC cleans them up in two steps:

1. Merge segments of the same speaker type separated by a gap shorter than 0.1 s.
2. Drop segments shorter than 0.1 s, which are most likely errors.

The cleaned output suits most analyses. The raw output is useful for debugging unexpected results, writing your own post-processing, or studying the model itself.

Keeping the raw output¶

Add the --keep_raw flag. The raw segments are then saved next to the cleaned ones:

<output_folder>/
├── 📂 rttm/          # Cleaned-up segments (default output)
├── 📄 rttm.csv
├── 📂 raw_rttm/      # Raw model segments (only with --keep_raw)
└── 📄 raw_rttm.csv   # (only with --keep_raw)

raw_rttm/ uses the same RTTM format as rttm/, with more and shorter segments.

Technical Reference¶

Processing steps applied to the model output:

1. Sliding-window averaging. The model runs on overlapping windows (step set by --stride_pct, default 0.25 of the window length), and the per-frame scores of overlapping windows are averaged.
2. Per-frame binarization of the four class scores with the active threshold set: f1.toml by default, hp.toml with --high-precision, or a custom file selected with --thresholds-location and --thresholds-name. See Threshold selection.
3. Gap filling: same-class segments separated by less than --min_duration_off_s seconds (default 0.1) are merged.
4. Short-segment removal: segments shorter than --min_duration_on_s seconds (default 0.1) are dropped.

Steps 3 and 4 use pyannote.audio annotation operations.

--keep_raw writes the segments obtained after step 2 to <output>/raw_rttm/. --save_probs writes the averaged per-frame probabilities from step 1 to <output>/probs/<file>-probs_dict_t.pt, a PyTorch file with one tensor per class. These files can be large.