recieve still resolves. It just ranks below receive.

Most of a merged dictionary is noise. One rule ranks it down.

A row with none of the four is weak. Weak rows are not deleted: exact lookup still finds them and says the word is probably misspelled. Search ranking and the spell-correction pool order on the flag rather than filter on it, so a weak row sits below the vouched-for ones, which is not the same as below every real word: a real word no source rated is weak too.

Counted from the built database. No native modules: SQLite is Node’s own node:sqlite, not better-sqlite3, and FTS5, the trigram tokenizer and custom functions all work through it, so the app installs on a machine with no MSVC build tools, which is why it was chosen.

Most of a caption’s delay is the speaker’s pause.

Lexica — live captions, recorded by the self-test

The italic line marked ··· is provisional, Over many episodes it learned., and a few frames later the accurate pass replaces it with the whole sentence. With eight glossary terms set up for this lecture the finished lines read 强化学习, 智能体观察了状态, 经验回放缓冲; the same audio without a glossary came out as 加强学习, 特工观察了国家, 重播缓冲. Synthetic speech fed at 4× and captured every 450 ms, so this shows the pipeline, not real-time pacing.

None of these is a benchmark. 0.58×, 1.50× and 4.00× compare runtimes on 5-second chunks of base.en. 9.1× and the WER are one 127-word clip synthesized from a lecture script, so there is a reference to score against, cleaner than a classroom; the same runtime gave 3.7× the day the table was first measured, and every WER came back identical. Throughput is not latency: 9.1× says audio is processed faster than it arrives. Splitting is on pauses, with a threshold that follows the room’s noise floor, never on a fixed clock that cuts sentences in half. The full table is in the Chinese README, and npm run eval:asr scores it again on the same clip.

The failures that read as working.

1. 01A common word, under a page of misspellingsECDICT ships recieve and wierd as entries. The weak flag now ranks them below every vouched-for row, and spell correction draws its candidates only from those.
2. 02Whisper repeating one sentence forever--audio-ctx 512 covers about 10.5 s. Fed 14 s, the decoder did not drop the tail; it looped, with no error. The window is now computed from the chunk limit.
3. 0363.8 became 638The local translator dropped a decimal point: 从 71.2 降低至 638. That is why there is no translation-only layout: every sentence sits beside its English.
4. 04A 429 that was not rate limitingElectron’s main-process fetch got 429 where net.fetch got 200 for the same URL. The check passed anyway: the fallback model happened to be right. It now asserts which channel answered.
5. 05反向传播反向传播A glossary rendering occurred twice where the English had the term once, and both were replaced. A rendering that occurs more often than its term is now left alone.
6. 06“Q-learning is on-policy”on-policy and policy both probed as 政策, and pinning a vague 政策性的 on on-policy made a true sentence false. A shared rendering now goes to the most general term.
7. 07Screenshots one step behindcapturePage() returns the last frame presented, and a covered window stops presenting. Minimized, 7–8 of 12 captures were stale and only 2–5 matched the one before byte for byte. Each capture now waits on two animation frames, or is not saved.
8. 08The Android copy driftedAndroid held a hand copy of the business layer: its word lists had no drill scope and its export lacked four columns. Both now load app-core.js, and a test checks its channels against the preload.

One source tree, and what Android leaves out on purpose.

The bridge detail that cost the most time: SQLiteDatabase.execSQL() refuses any statement that returns rows, several PRAGMAs do in their assignment form, and the error never mentions PRAGMA. The note came out of its first launch on a real phone.

A dictionary is mostly ranking.

Every picture is from a real run.

$ npm run shot
[shot] ✓ 滚动容器正常（#stage）
[shot] ✓ 翻译页双语对照 2 句
[shot] ✓ 译文里用上了术语表的译名：策略、经验回放缓冲
[shot] ✓ 对照组未被改动
[shot] ✓ 水平检测出结果：答对 3 / 30 · 10% · 3 个分层
…
38 张，退出码 0

$ npm test
ℹ tests 270
ℹ pass 270
ℹ fail 0

One download, then no network.

$ npm install
$ npm run data    # ~800 MB of corpora → data/dict.db
$ npm start
$ npm test        # 270; on a fresh clone, those that
                  # need data/dict.db skip

# captions and translation are separate downloads
$ npm run fetch:asr     # whisper.cpp + base model, ~90 MB
$ npm run fetch:model   # opus-mt-en-zh, 118 MB