Release 2.0.0 · ethnosdev/isochron

2.0.0 - 2026.10.21

Neural Forced Alignment: Replaced legacy macOS say TTS, acoustic MFCC extraction, and Dynamic Time Warping (DTW) with Meta's MMS-300M Wav2Vec2 + CTC acoustic model.
1,130+ Languages Supported: Integrated uroman: ^1.1.0 for universal phonetic romanization and phonetic normalization across scripts (Greek, Hebrew, Cyrillic, Arabic, Devanagari, CJK, etc.) without requiring per-language dictionaries.
Sub-10ms Acoustic Precision: Combines 50 Hz CTC Viterbi decoding with 10 ms adaptive energy envelope pause detection, achieving a 45 ms aggregate median error and 93.6% accuracy within 200 ms across multilingual evaluation fixtures (100% in Hebrew, 96% in Greek).
Pure-Dart ONNX Runtime FFI: Built direct dart:ffi bindings to the native ONNX Runtime C API with dynamic int8 quantization (model_int8.onnx, 302.6 MB), executing offline on Apple Silicon CPU SIMD hardware.
Model Status Panel: In-app management panel to download, inspect, verify SHA-256 integrity, cancel, or delete local ONNX models.

Adaptive Silence Thresholding: Analyzes frame RMS energy in dBFS with $P_{15}$ noise floor and $P_{80}$ active speech tracking to adapt to variable recording conditions.
Simple Pause Snapping: Snaps verse boundaries directly into preceding acoustic breath pauses with safety margin clamping and calibrated 190 ms lead-in.
Unread Omission Detection (find_unread): Automatically detects skipped verses where reading rate exceeds physical human limits (> 25 chars/s) and acoustic probability drops, marking them with -1.0 timestamps and omitted flags.

Verse Confidence Scores: Generates calibrated confidence ratings (0–100%) blending acoustic character probabilities with boundary silence gap placement.
Diagnostic Anomaly Flags: Automatically flags verses with omitted, not_at_pause, poor_text_match, long_verse, or short_verse to accelerate quality assurance.
Keyboard Flag Navigation: Press F to immediately jump to the next verse with diagnostic flags (with auto-wrap).
Amber Warning Visuals: Flagged verse boundaries and waveform regions are highlighted in amber to draw immediate visual attention during review.

Hard Constraint Trellis Solving: Re-aligning a track preserves user-edited and reviewed boundaries as strict hard constraints, solving the CTC trellis only for intermediate unreviewed verses.
Re-alignment Safety: Safe confirmation dialog offers "Keep My Edits", "Overwrite All", or "Cancel", and automatically saves .bak backup files prior to re-alignment.

Keyboard-Driven Review: Nudge active boundaries forward or backward at 10 ms (Cmd+Opt+Arrows), 50 ms (Cmd+Arrows), or 250 ms (Cmd+Shift+Arrows) with instant auto-replay from 0.5s before the boundary.
Boundary Snippet Audition: Press B to play 1.5s before and after the active boundary.
Review and Advance: Press R to mark a verse as reviewed and select the next verse.
Ghost Markers: Visual dashed lines display the original AI recommendation when a boundary has been adjusted.

CLI Forced Alignment: Direct headless alignment via isochron CLI with --lang, --lead-in, --star-score, and --snap-pauses options.
Built-in eval Command: Scores alignment accuracy against ground-truth fixtures with full metrics reporting (Median, Mean, p90, p95, % <= 100/200/500ms, bias) and CSV export.
Built-in model Command: CLI subcommands model status, model download, and model delete.

Removed legacy macOS say TTS synthesis.
Removed Pins service, -pins.json sidecar files, and manual pin toggling (superseded by "Keep My Edits" constrained alignment).
Removed Team Collaboration ClaimService and claim locks.
Removed transliteration dictionaries (superseded by uroman).