Filler removal processor

[dinooo13]

What

New FillerRemover TextProcessor, first in the pipeline (fillers → dictionary → whitespace). Strips hesitation sounds with a single case-insensitive regex pass; a settings toggle shows up in the Processing tab automatically.

Coverage

Conservative by design — only particles that are never content words in English, German or Spanish:

• English: "uh", "um", "erm", "mm", "hmm"
• German: "äh", "ähm", "öh", "öhm", "mh"
• Spanish: "eh", "em"
• Elongations via repeating-letter patterns("uhhh", "ähmm", "mmm")
• None of the judgment calls: "like", "you know", "pues", "eigentlich" stay untouched

Commas and trailing punctuation travel with the filler ("I, um, think" → "I think"); a leading filler carries its capitalisation over ("um the meeting…" → "The meeting…").

Critical-path benchmark

M1, same session and conditions for both runs (machine under load ~4.3–4.5, thermal tag "fair" on both — this is a like-for-like comparison, not a clean idle run). pladder-cli bench times engine.transcribe; processors sit after it, so the two tables are identical within spread:

Before:

Fixture	Audio	Engine (median)	Spread	WER
10s	9.7 s	0.212 s	16 %	0.0 %
30s	31.8 s	0.413 s	8 %	0.0 %
60s	60.8 s	0.616 s	7 %	1.4 %
2m	125.1 s	1.070 s	11 %	0.9 %
5m	315.6 s	2.331 s	4 %	0.6 %
10m	631.4 s	4.509 s	4 %	0.6 %

After:

Fixture	Audio	Engine (median)	Spread	WER
10s	9.7 s	0.207 s	22 %	0.0 %
30s	31.8 s	0.417 s	3 %	0.0 %
60s	60.8 s	0.620 s	3 %	1.4 %
2m	125.1 s	1.072 s	41 %	0.9 %
5m	315.6 s	2.542 s	28 %	0.6 %
10m	631.4 s	5.499 s	13 %	0.6 %

Processor stage measured directly (release build, per process() call): ~400 µs worst case on the 2,196-word 10-minute reference text, ~830 µs on a fillers-dense variant — two orders of magnitude below the noise line of both runs.