Alignment decisions

Architecture: alignment. Related: SGPA.

2026-10-10 — Own CTC alignment instead of torchaudio

Decision: alignment.ctc.forced_align is a NumPy Viterbi over the blank-extended target sequence, resampling uses librosa, and the CTC model is a CtcModel protocol that Wav2Vec2Ctc implements. The audio extra holds only librosa and soundfile.

Why: The earlier port called torchaudio.functional.forced_align, which torchaudio deprecated for removal from its application programming interface (API), and torchaudio releases trail and pin torch versions (2.11 against our torch 2.13).

Consequences: Alignment, boundary refinement and audio I/O are tested in continuous integration with a known-truth fake model; only Wav2Vec2Ctc needs the transformers extra. Rejected: keeping torchaudio (deprecated API, version lock); MagicMock-patched tests of the model (they test the mock, not the alignment).