0.3.0

Added

• A second downloadable voice: English (Piper LJSpeech, en_US, medium, dataset public domain)
  next to the German Thorsten. A new setting Voice for export switches between them; it also
  applies to reading aloud with the downloaded voice.
• On first start the voice follows Obsidian's display language (English UI → English voice). A
  choice you have made is never overwritten.

Changed

• Worker and ONNX runtime are now shared between voices: a second voice downloads only its
  model (≈ 60 MB instead of ≈ 75 MB), the voice list shows what each one still costs, and
  Remove keeps the shared runtime as long as another voice needs it.
• The synthesis worker carries both eSpeak-NG dictionaries (English and German), ≈ 2.8 MB instead
  of ≈ 2.2 MB; only the one the selected voice needs is loaded at runtime.

0.2.0

Changed

• Factory tempo of the Piper voice is now 0.85× (length_scale) for export and for reading aloud
  with the downloaded voice — chosen by listening test; the Rate setting still scales relative to it.

0.1.1

Fixed

• The voice download from the GitHub release failed inside Obsidian (Failed to fetch): the
  renderer's fetch is blocked by CORS on GitHub's redirect. Downloads now go through Obsidian's
  requestUrl (per-file timeout 15 min). Progress advances per file rather than per byte, and
  cancelling takes effect between files.
• Cache keys of the downloaded assets no longer depend on the download URL (version + file name),
  so a different source (mirror, local server) never hides an existing download.

Changed

• Settings: the engine row says “Ready · v0.1.0” instead of repeating the size; the decimal
  separator of sizes follows the UI language.

0.1.0

Added

• Read notes or the current selection aloud with the system voices (speechSynthesis):
  start, pause/resume, stop, voice and rate settings, status bar item, ribbon icon.
• WAV export as an opt-in: enable it in the settings, then press Download to fetch the
  German Piper voice (de_DE-thorsten-medium) plus its runtime (ONNX Runtime WASM, eSpeak-NG
  phonemizer) from this repository's GitHub release — nothing is downloaded before that click.
  Assets are verified against SHA-256 checksums embedded in main.js, kept in the browser
  cache outside the vault, and can be removed again from the settings.
• Output profiles: phone system (8 kHz mono 16 bit, e.g. for 3CX mailboxes) or the voice's
  native 22.05 kHz; target folder, file-name pattern ({{note}}, {{date}}), optional link
  insertion; files are never overwritten (-2, -3 suffixes).
• Optionally read aloud with the downloaded voice instead of the system voice.
• Markdown is turned into speech text before synthesis: frontmatter, code blocks, embeds and
  comments are dropped; links speak their display text; headings, lists, tables and callouts
  become chunks with natural pauses.