Uncensored

Uncensored restores words hidden as [__] in YouTube captions, primarily for deaf and hard-of-hearing viewers. Audio, captions, and inference stay local.

Modes

The popup slider has five stops (default: Rules first):

Mode	Behavior	CPU load
Off	Leave captions unchanged.	None
Rules only	Fill from caption context; uncertain slots abstain.	Minimal
Rules first	Rules fill what they can; Whisper scores only unfilled or ambiguous slots.	Moderate
Whisper first	Rules fill instantly; Whisper checks every slot and replaces the fill.	High
Whisper only	Local Whisper only.	High

Settings from versions with two switches map to the matching mode (both on → Rules first). The evaluator's rules+whisper mode is "Whisper first".

Current evaluation metrics

The canonical snapshot is docs/evaluation-metrics.json, as of 2026-10-02.

Tier	Status	Scored slots	Precision	Coverage
Manual-auto	development	26,429	89.60%	44.81%
Synthetic	development	149,135	89.94%	40.16%
Auto-auto	development	10	100.00%	80.00%
Historical validation (unverified)	diagnostic	9	100.00%	11.11%
Held-out test	held-out-test	—	—	—
Any-candidate oracle	diagnostic	—	90.81%	49.83%
Synthetic Filmot gap sample, rules-only	diagnostic (stale/superseded)	110	85.37%	31.82%
Manual Filmot wave, rules-only	diagnostic (stale/superseded)	275	69.72%	27.64%
Dense audio set, rules only	diagnostic	2,493	95.18%	49.10%
Dense audio set, Whisper	diagnostic	2,449	96.94%	80.16%
Dense audio set, Whisper first	diagnostic	2,449	96.55%	86.85%
Dense audio set, rules first (default)	diagnostic	2,449	95.41%	84.89%

Development rows are in-sample. No held-out test score is available. Rows marked stale/superseded are historical diagnostics, not current-code benchmark results. Current evaluator fingerprints: rules 3795:39:11g0z2u, aux 8hsabx, engine hof9ii.

The later frozen audio comparison is documented separately as a historical diagnostic: its reports lack the creator-manifest provenance required for canonical held-out metrics.

Evaluation caveats

This is an English, playlist-selected development corpus. Videos qualify only when separate automatic and manual English captions exist, the automatic track contains [__], and the manual track supplies a supported swear. It therefore does not represent YouTube generally or the natural prevalence of censored captions.

Rules were tuned and measured on this same corpus, so rules-only results are in-sample development figures rather than held-out estimates. Metrics are micro-averaged per caption slot, which gives prolific creators more influence. Evaluator reports also include summary.creatorMacro and summary.creatorMacroByPairClass, which weight each identified creator equally. The archived replay's roughly 223k slots include both labeled and unlabeled slots; accuracy is computed from labeled slots only, with unlabeled slots reported separately. Manual caption omissions, paraphrases, and timing differences can still introduce alignment error.

Architecture

• page-hook.js observes YouTube JSON3 captions and SABR media responses.
• timedtext.js parses captions and gives every [__] slot a stable identity.
• rules-compiler.js validates and expands reusable rule declarations.
• rule-data/ separates language sets, exact rules, and grammar; rules-data.js assembles them with compact candidate priors, and rules.js supplies replacements and Whisper candidates.
• sabr-parser.js extracts audio already downloaded for playback.
• audio-capture.js decodes only token-adjacent audio, queues inference, and reapplies results when YouTube redraws its rolling caption rows.
• background.js, offscreen.js, and whisper-module-worker.js keep Whisper off the page thread. Chromium uses an offscreen document; Firefox uses its background page.

Caption, audio, and DOM state are scoped by video, track, and navigation generation. In hybrid mode, deterministic rules provide immediate provisional fills while full-vocabulary Whisper checks every slot. A Whisper result replaces the fill; if Whisper abstains, the provisional rule remains for coverage. Rules-only mode continues to apply the full deterministic rule set. Decoding and the model stop when captions contain no censored slots. Decoded segments are discarded when no pending token needs them. Transport/inference failures get at most one retry after 250 ms; valid empty Whisper decisions remain abstentions. Superseded navigation, track, seek, or mode requests do not retry or apply stale results. A host timeout does not cancel inference already running in the worker.

The local quantized whisper-tiny.en model runs through vendored Transformers.js and ONNX Runtime Web. No remote code is loaded. Whisper does not transcribe freely. It encodes one 30 s window starting 3 s before the next upcoming slot and scores every slot inside it: each allowed word is scored as the continuation of the caption words spoken since the window start. It fills only when the best word's log probability is at least -4, and uses the next caption word to choose among forms such as fuck/fucking. Upcoming slots are scored first so words are usually ready before their captions appear.

Word vocabularies

ALLOWED_WORDS is the broad set of supported censored words accepted from local Whisper. RULE_WORDS is its conservative subset that deterministic context rules may emit. A word can therefore be recognized from audio without becoming a context-only guess. WORD_ROLES is a broader authoring catalog and does not by itself permit runtime output.

Add newly supported censored words to ALLOWED_WORDS only after the manual-auto audit confirms repeated whole-word evidence absent from visible automatic captions. A split form such as chicken [__] does not validate chickenshit. NOT_CENSORED_WORDS records swear labels that YouTube leaves visible in those tracks and is disjoint from ALLOWED_WORDS; the vocabulary-discovery audit skips them so a nearby visible swear cannot become a false new label. Promote one to RULE_WORDS only when an exact rule or grammatical frame can identify it at the required precision. List order is stable because Whisper uses it to break ties.

Future plans

• Continue replacing broad wildcard rules with validated concrete alternatives.
• Continue validating creator-diverse paired-caption rules on held-out videos.
• Measure in-browser Whisper scoring latency and recalibrate its gate on new creator-diverse audio labels.
• Re-run the caption audit after vocabulary changes; keep visible automatic labels in NOT_CENSORED_WORDS and regenerate synthetic fixtures deliberately.
• Follow docs/SESSION_HANDOFF.md for current state and next steps; rule evidence gates are in docs/RULES.md.

Limitations

• Local inference can finish after a caption scrolls away on slow hardware.
• Ambiguous visible rows are left unchanged rather than guessed.
• Seeking to discarded audio requires YouTube to fetch that media again.
• Only English is currently supported.

Enable debug logs in the YouTube tab, reload, and show Verbose messages:

localStorage.setItem("uncensoredDebug", "1")

Development

Whisper-only work should not change deterministic patterns unless rule behavior is explicitly in scope.

Run tests, both builds, ZIP validation, and the archived corpus regression. When its ignored local source reports are present, this replays at least 200,000 unique slots; the current inventory is 223,478:

npm test

Before release, also require metric provenance and run pinned Firefox lint:

npm run test:release

A clean checkout skips only the archived replay until those local reports are restored; its unit test still runs from temporary miniature reports.

The release check validates the canonical metrics snapshot and generated README table. After an intentional handoff update, regenerate that table with:

node tools/evaluation-metrics.js --write-readme

Optional checks:

npm test -- --benchmark
npm test -- --browsers --chromium-only --workspace URL [URL...]
npm test -- --all --chromium-only --workspace URL [URL...]

The native, headless Firefox smoke checks one URL only (no SPA, pause, or playlist coverage). It defaults to the an5 video, whisper-first, and 90 s:

npm run test:firefox
# Recommended strict pure-audio check:
npm run test:firefox -- --mode=whisper --rate=1 --seek=20:45 --until=90 \
  --expect=fucking 'https://www.youtube.com/watch?v=an5iFYcjWUM'

Expected words are optional except in rules mode. See Firefox smoke details for dependencies, isolation, evidence, and artifacts. This is a runtime diagnostic, not accuracy validation of every mode.

Chromium separately covers multiple SPA redirects (requires network and Chromium; keep --workspace or use safe headless execution):

node tools/browser-smoke.js --chromium-only --workspace --headless --via=search --mode=hybrid \
  'https://www.youtube.com/watch?v=kTeQSzHGWyw' \
  'https://www.youtube.com/watch?v=an5iFYcjWUM' \
  'https://www.youtube.com/watch?v=jNQXAC9IVRw'

Both runners accept off, rules, rules-first, whisper-first, and whisper; legacy aliases both-off, rules-only, hybrid, and whisper-only remain. Chromium also accepts --via=search|direct|home, --auto-next=N, --until=SECONDS, --pause=SECONDS, and --verbose. The navigation-expression unit test runs without launching browsers.

Specialized corpus acquisition, benchmark evaluation, audio annotation, and rule-mining workflows live in docs/DEVELOPMENT.md. Provenance details remain in docs/CAPTION_CORPUS_AUDIT.md; the auxiliary text-source notes are in docs/OPENSUBTITLES_EXPANDED_AUX.md and docs/SBCSAE_CORPUS.md.

Build

./build.sh 1.6.1

This creates separate Chromium and Firefox ZIPs in dist/. See AMO_SOURCE.md for submission and vendored-runtime notes.