WordAlign 2.2.0

Bad audio. A new robust flow (--profile robust, GUI preset Robust) measures the recording, cleans only the parts that need it, lets several ASR engines vote, times words with forced alignment on the original audio, and sorts every cue into a review queue. Every new stage is optional and skipped with a message when its model is not installed. English first; see docs/robust_flow.md.

Upgrade notes (behaviour changes)

• Word timing uses anchors + forced alignment by default (--timing anchor): agreeing engines give anchors, Qwen3-ForcedAligner (or wav2vec2) aligns the windows between them on the original audio, MFA and pause-aware interpolation fill the rest. --timing fill_only brings back the 2.1 waterfall; --no-forced-align keeps the engine times.
• Every run with anchor timing (the default) writes <name>_review.json (cue tiers Accept / Quick check / Redo with reasons) unless --no-triage is given.
• The audio analysis stage runs by default when its model is present (Silero VAD; --no-vad turns it off), and the hallucination filter drops engine words the audio does not support (--no-hallucination-filter).
• Speaker labels use pyannote Community-1 (falls back to 3.1); accept its terms on Hugging Face once.
• Transcript-free mode skips Vosk when two timed voters already anchor the timing and a forced aligner (Qwen3-ForcedAligner or wav2vec2) is installed.
• The Qwen3-ForcedAligner language list in the model catalog was wrong (it listed Arabic); it now matches the model: zh, en, yue, fr, de, it, ja, ko, pt, ru, es.

Added

• Audio analysis: Silero VAD (ONNX, no torch in the app), VAD-planned ASR chunks, WADA-SNR, narrowband and clipping detection, and timed sound events (PretrainedSED BEATs) per chunk.
• Gated cleanup (--cleanup auto|all|off): RoFormer vocal separation for loud music, RoFormer crowd removal, MossFormerGAN for steady noise, mixed into a separate ASR track; never on phone or clipped audio.
• ROVER voting over Cohere Transcribe, Qwen3-ASR, parakeet-ultra, WhisperX and Vosk (--voters), per VAD chunk, with each word's agreement and the losing readings kept for review.
• Overlap map (--overlap-map) and overlapped-speech ASR with SE-DiCoW (--overlap-asr); two-speaker dash cues (--dash-cues, --dash-style).
• SDH tag proposals from the sound events (--tags).
• Review queue in the GUI: loop a cue, apply a voter's alternative with one click, accept with Ctrl+Enter. Corrections go to <name>_corrected.srt (plus an edit log); the pipeline's files are never changed.
• Triage calibration: wordalign-bench fit-triage --jobs FOLDER learns the risk weights from your corrected jobs.
• Benchmark: wordalign-bench builds a suite from public data (LibriSpeech alignments, MUSAN noise, FMA music) and scores WER, timing, invented words and triage per condition; see docs/benchmark.md.
• Setup: scripts/install_robust.ps1 (venvs and models), scripts/fetch_models.py (--list shows what is installed), wordalign --doctor, scripts/smoke_robust.py (one real call per backend).
• Configurable locations: --models-dir / WORDALIGN_MODELS_DIR, WORDALIGN_VENVS_DIR, WORDALIGN_<NAME>_PYTHON for every backend, WORDALIGN_VAD_MODEL, WORDALIGN_SED_DIR.
• The model catalog lists the new models with their licenses.

Measured (wordalign-bench, quick suite, RTX 4070 Ti)

Five minutes of LibriSpeech speech in nine conditions (clean, noise 5 dB, music 0 dB, six-talker babble 0 dB, reverb 1.0 s, μ-law phone, a second talker at −5 dB, music only, noise only), every run with an empty cache. "2.1" is the 2.1 behaviour run by the 2.2 code (--timing fill_only --no-vad --no-hallucination-filter --no-triage).

	2.1	2.2 defaults	--profile robust
Word-start error, known transcript (mean)	868 ms	69 ms	70 ms
Words within 100 ms, known transcript	83.0%	90.1%	90.1%
WER, no transcript	16.9%	15.0%	15.3%
Word-start error, no transcript (mean)	225 ms	73 ms	106 ms
Invented words per 10 min of music/noise only	300	0	0
Erroneous cues the review queue flags (no transcript)	–	88%	77%
Run time for the 5-minute file	238–289 s	234–282 s	195–268 s

By condition: reverb WER 43% → 2.6%; babble timing with a known transcript 5.8 s → 0.2 s. Worse than 2.1: transcript-free WER under 0 dB music with sung vocals (0% → 12–13%: the voters also transcribe the lyrics) and under 0 dB babble (73% → 84–87%, unusable either way). On this synthetic suite the robust profile's extra voters and cleanup did not beat the 2.2 defaults; whether they help on real recordings with music beds, crowds and crosstalk is not measured yet, so compare on your own material (see docs/benchmark.md).

Fixed

• Qwen chunks that ran out of GPU memory are split and retried; pieces that still fail are reported instead of being silently empty.
• WhisperX finds its venv through the same lookup as the other backends.
• MFA from a conda environment (the bundled venvs\mfa, WORDALIGN_MFA, or aligner/mfa under your conda folders) is now preferred over an mfa on PATH, and is run as if activated: its DLL folders on PATH and the user site-packages ignored. A pip-installed MFA in the user site-packages (no kalpy on Windows) used to win and fail on every run.
• --doctor starts MFA the same way instead of only finding it, so a broken MFA is reported.
• MFA only fills words that have no time; a word already timed by an engine, the aligner or a voter is no longer re-aligned (and overwritten). Each gap is aligned with three timed words of context on each side, which fixed gap words being pinned to the start of their slice (0.5 s early).
• A forced-aligned word no longer vanishes from the word-level SRT when the aligner gives it no length.