feat: anti-repetition protection for whisper transcription

[egsok]

Problem

Whisper (especially large-v3) is prone to hallucination loops — repeating the same phrase indefinitely, particularly during silence or at audio segment boundaries.

The root cause: Whisper uses the previous segment's output as context for the next. A repetition in segment N propagates to N+1, N+2, etc. — creating an exponential feedback loop. This is one of the most widely reported Whisper issues:

• openai/whisper#679 — "A possible solution to Whisper hallucination" (37 comments, 147 replies) — foundational discussion on the root cause
• whisper.cpp#1490 — "Large Model hallucination and repeating", maintainer recommends -mc + large-v2
• whisper.cpp#1507 — maintainer recommends --max-context 64 or 32, --beam-size 5, --entropy-thold ~2.8
• whisper.cpp#1244 — confirms -mc 0 = condition_on_previous_text=False

Solution

Three layers of protection:

1. --max-context 128 — limits cross-segment context to 128 tokens, breaking the feedback loop. Does NOT affect initial_prompt (custom prompt and dictionary words are always sent in full). Maintainer recommends 64 (#1507); we use a more conservative 128 to preserve cross-segment continuity, relying on the regex dedup layer to catch any loops that slip through.

2. --entropy-thold 2.8 — discards segments where model confidence is very low (high entropy = silence/noise/hallucination). Default is 2.4; raised to 2.8 per maintainer recommendation in #1507. Zero performance overhead.

3. Regex dedup — post-processing safety net that removes consecutive repeated phrases (10+ chars) from final output. 10-char minimum avoids false positives on common short phrases.

Parameter	Value	Default	Rationale
--max-context	128	224	Conservative vs maintainer's 64 — regex dedup catches the rest
--entropy-thold	2.8	2.4	Per maintainer recommendation (#1507)
Regex min length	10 chars	—	Avoids false positives on "I think", "you know", etc.

How Whisper context works

initial_prompt (custom prompt + dictionary)
  → always used in full (up to 224 tokens)
  → NOT affected by --max-context

cross-segment context (previous segment output)
  → fed back to next segment for continuity
  → --max-context limits THIS to prevent loops

UI

New "Suppress repeated phrases" toggle in Settings → Transcription (default: on). When disabled, only regex dedup is skipped — whisper.cpp flags remain active as they improve quality regardless.

Files changed

File	Change
src/helpers/whisperServer.js	--max-context 128 and --entropy-thold 2.8 flags
src/helpers/audioManager.js	removeRepetitions() regex dedup + wiring in processTranscription()
src/components/SettingsPage.tsx	Toggle UI in Transcription section
src/hooks/useSettings.ts + src/stores/settingsStore.ts	suppressRepetition boolean setting
src/locales/*/translation.json	i18n for all 10 locales

Test plan

• Record 30+ seconds with pauses — verify no hallucination loops
• Record normal speech — verify transcription quality unchanged
• Verify custom prompt and dictionary still work with --max-context 128
• Toggle off → verify repeated text passes through
• Check debug logs for --max-context 128 and --entropy-thold 2.8 in server startup args