[harshmathurx]

Lector now reads in whole sentences and says numbers the way a person would.

Why

• Chunking was cutting phrases. Chrome's sentence splitter never breaks before a lowercase word, so all-lowercase posts (common on X) became one giant "sentence". The chunker then cut it at arbitrary words, e.g. "…the more exposure | i got talking…". Kokoro restarts its intonation at every cut, which is what sounded off.
• Numbers and symbols were read literally. "$2.5M" was read as "dollar two point five em".

No LLM. Lector doesn't load one. Adding one would cost a 350–500 MB download, seconds to minutes before the first word, and a model that can rewrite words, which breaks the page highlighting. Rules are instant and testable.

What changed

• Chunker (shared/chunker.ts)
  • Its own sentence pass handles lowercase text, using sbd's abbreviation list (MIT), initials and 1/ thread numbering.
  • Segments are whole sentences up to 420 chars; short neighbours are joined while they stay under 140 chars.
  • Over-long sentences break at a scored point: ; : — first, then commas, then clause words, then a space. It avoids short leftover fragments and never breaks inside a number, currency amount or unit.
  • Pauses depend on the punctuation (?/!, ., : ; —, ,, or none).
  • The short first segment for a fast start is kept.
• Normalizer (shared/normalize.ts), rules ported from misaki (Apache-2.0), plus number-to-words (MIT, about 1.3 KB). It covers:
  • currency with magnitudes ($ £ € ₹ ¥, k/M/B/bn/cr/lakh), including "forty billion dollar valuation"
  • percentages and percent ranges, YoY and similar
  • ranges (phone numbers are left alone), ordinals, fractions, multipliers
  • units, with "5m" read as metres or million by context
  • ISO and slash dates, and times
  • Roman numerals in context ("World War II", but "Type C" stays a letter)
  • initialisms Kokoro mangles, versions, @Handles and #hashtags, X 1/ numbering
  • Output is checked to stay stable under kokoro-js's own normalizer.
• Engine guard (engine.worker.ts): a segment over 500 phoneme tokens is split and its audio joined. kokoro-js used to silently cut off the end.
• Extraction (content.ts): Draft.js blocks (X Articles) are treated as paragraphs. When Readability's copy of a block maps to a live wrapper, the live paragraphs are read instead. Output on paulgraham.com and Wikipedia is identical to main.
• Report a problem in Settings: opens a pre-filled GitHub issue containing only the Lector and Chrome versions, OS, engine, voice quality, voice, speed and last error. It never includes the page address or article text, and needs no new permissions.
• Version 0.4.0; README, CLAUDE.md and THIRD_PARTY_NOTICES updated.

Verification

• tsc is clean and bun test passes (205 tests). Your screenshot paragraphs are fixtures, and each review finding has a test.

• scripts/e2e.mjs with real Kokoro passes every check on WebGPU and NO_GPU at the final tip.

• That run caught a release blocker the unit tests couldn't: the guard called .call on the tokenizer, which doesn't have one. Fixed in 8d5b337.

• The guard fired on a real page (726 tokens) and the full 41.8 s of audio was produced, so nothing was truncated.

• Bench, warm runs, one run each, so a rough comparison only:

  Backend	Time to first audio, main → this branch	Speed (RTF)	Stalls
  WebGPU	2.44 s → 1.46 s	unchanged	0
  WASM	5.48 s → 2.20 s	unchanged	0

• axe finds zero violations on the Settings sheet.

• A high-effort code review found 10 issues, all fixed.

Needs a human

• Listening by ear. Old and new audio for the screenshot paragraphs and a number-heavy paragraph is in the session scratchpad (verify/ab/).
• A real X Article. The extraction fix was only tested on fixtures, because X needs a login.
• Phonemes not checked. "ay M" for a.m. wasn't checked at the phoneme level.

🤖 Generated with Claude Code