Skip to content

ROADMAP.md

Latest commit

 

History

History
130 lines (101 loc) · 5.52 KB

ROADMAP.md

File metadata and controls

130 lines (101 loc) · 5.52 KB

Reed Roadmap

Phase 1 — Interactive Enhancements

1.1 Command Autocomplete (Tab Completion)

Effort: Small · Priority: High Status: Partial — prompt_toolkit is already in use with AutoSuggestFromHistory.

  • Add a custom Completer to _make_prompt_session() that completes /quit, /exit, /help, /clear, /replay (and future commands like /pause, /stop)
  • Fuzzy-match support so /h → /help
  • Extend to file paths for a future /open command

1.2 Non-blocking Playback Controller ✅

Effort: Medium · Priority: High Dependencies: None Status: Done

  • ✅ Create PlaybackController class with background thread playback
  • ✅ Pause/resume via SIGSTOP/SIGCONT (Unix)
  • ✅ Stop via proc.terminate()
  • ✅ Thread-safe state management with threading.Lock
  • ✅ Integrated into interactive mode
  • ✅ Zero new dependencies (stdlib threading + signal)

1.3 Playback Controls — Pause, Play, Stop

Effort: Small · Priority: High Dependencies: 1.2 (non-blocking playback) ✅ Status: Ready to implement

  • Add interactive commands: /pause, /play (resume), /stop
  • Wire commands to PlaybackController.pause(), .resume(), .stop()
  • Display playback state in prompt (e.g., ▶ Playing…, ⏸ Paused)
  • Add to autocomplete and /help output

Phase 2 — File Format Support

2.1 EPUB File Reading ✅

Effort: Medium · Priority: High Dependencies: None Status: Done

  • ✅ Create _iter_epub_chapters() yielding (chapter_number, total_chapters, text) — mirrors _iter_pdf_pages()
  • ✅ Wire into main() with .epub suffix detection alongside the existing .pdf path
  • ✅ --pages flag selects chapters for EPUBs
  • ✅ Strip HTML tags from EPUB XHTML content (stdlib html.parser)
  • ✅ Navigation documents filtered out via spine ordering
  • ✅ Zero external dependencies — uses stdlib zipfile + xml.etree.ElementTree

Phase 3 — Reading Progress & Bookmarks

3.1 Save & Resume Reading Position

Effort: Medium · Priority: Medium Dependencies: Phase 1.2 (stop controls) ✅, Phase 2.1 (epub support desirable) ✅ Status: Ready to implement

  • Create a JSON bookmarks file at _data_dir() / "bookmarks.json"
  • Schema: 
    {
  "/abs/path/to/book.pdf": {
    "page": 12,
    "char_offset": 340,
    "timestamp": "2026-02-18T10:00:00"
  }
}
  • On /stop or Ctrl-C during file reading, persist current page + character offset
  • Add --resume / -r flag: reed -f book.pdf --resume picks up where it left off
  • Add reed bookmarks subcommand to list/clear saved positions
  • Interactive command: /bookmarks to list, /resume to continue last file

Phase 4 — Streaming Audio

4.1 Streaming TTS Playback

Effort: Large · Priority: Medium Dependencies: Phase 1.2 (playback controller) ✅ Status: Ready to implement

Currently reed generates the full WAV file before playing. Streaming removes that wait.

  • Use piper --output-raw which streams raw PCM to stdout
  • Pipe piper's stdout directly into the audio player's stdin (e.g., aplay -f S16_LE -r 22050 -c 1 on Linux, ffplay -f s16le -ar 22050 -ac 1 - on macOS/Windows)
  • On macOS: afplay doesn't support stdin; switch to ffplay or sox play for streaming, or use python-sounddevice to play raw PCM frames as they arrive
  • Chunk-based playback loop:
    1. Start piper Popen with stdout=PIPE
    2. Start audio player Popen with stdin=PIPE
    3. Read chunks from piper → write to player
    4. Integrate with PlaybackController for pause/stop mid-stream
  • Fallback: keep the current generate-then-play path when streaming isn't available

Phase 5 — Language Expansion

5.1 Lithuanian Language Support — Voice Model Training

Effort: Large · Priority: Medium Dependencies: None (can be started anytime, independent of reed code changes)

No Lithuanian piper voice model exists yet — this requires training one from scratch.

  • Dataset: source or record a Lithuanian speech dataset (e.g., from Common Voice, M-AILABS, or custom recordings); needs ~2–10 hours of single-speaker audio with transcripts
  • Preprocessing: segment audio, normalize text, generate phoneme alignments (espeak-ng has Lithuanian support: lt)
  • Training: use piper-recording-studio for data collection and piper training pipeline (VITS-based)
    • Train low, medium, and high quality variants
    • Export to .onnx format for piper runtime
  • Hosting: publish the trained model to HuggingFace (or self-host) so _model_url() can resolve it
  • Reed integration: once the model is hosted, add lt_LT-<voice>-medium as a downloadable voice; optionally add --lang lt shorthand
  • Validation: native speaker review for pronunciation quality

Summary Timeline

Phase Feature Effort Dependencies Status
1.1 Command autocomplete Small — Partial
1.2 Non-blocking playback controller Medium — ✅ Done
1.3 Pause / play / stop commands Small 1.2 Ready
2.1 EPUB reading Medium — ✅ Done
3.1 Save & resume position Medium 1.2, 2.1 Ready
4.1 Streaming audio Large 1.2 Ready
5.1 Lithuanian voice model training Large — —

Phase 1.2 is complete ✅. Phases 1.3, 3.1, and 4.1 now have their dependencies satisfied and are ready to implement.