How does CTC forced alignment turn a transcript into word timestamps (frames, blanks, merge, score), and is the torchaudio API still available?

Problem: you have an accurate transcript and need word start/end times for subtitles, dubbing timing or text-based editing. Conditions: CTC-trained acoustic model (tutorial uses the multilingual MMS alignment bundle). Method: compute the emission (frame-wise token probabilities); call forced_align with the tokenised transcript and blank index 0; the result is in the emission's frame coordinates with blanks and repeats; merge_tokens collapses repeats into token spans with scores; group tokens into words; word score is the average weighted by span length; convert frames to seconds with the ratio of waveform samples to emission frames. A special star token handles missing transcript. Failure modes: Inference: without the star token, untranscribed speech must be absorbed by neighbouring transcript tokens (the tutorial only says the star token handles missing transcript), and low word scores flag mismatched text. Limitations: the page itself warns that torchaudio moved to maintenance in 2.9 and that these APIs were deprecated in 2.8 and removed in 2.9, so pin an older torchaudio or use another aligner (WhisperX's README says it uses dynamic programming for alignment with language-specific models); the extension of this to dubbing sync tolerances is not in this source. Version/date: stable docs rendered with torchaudio 2.11.0a0, fetched 2026-10-24. Curator qualifications: All cited documentation excerpts located on the live pages by curator GET; no video sources. Correction folded 2026-10-24 from C3-B866-01 (curator-reviewed, C3R-0613/0616): Problem: a prior finding (C3-B844-10) asked whether the torchaudio CTC forced-alignment API is still available; the question premise 'removed in 2.9' is outdated. Conditions: applies to code using torchaudio.functional.forced_align. Method: the maintainers' plan (issue 3902) said APIs relying on C++ such as forced-alignment 'may be dropped' and would be removed in 2.9; a 2026-01-21 comment on the 2.10 release says forced_align, lfilter, RNNTLoss, CUCT and overdrive 'are staying' after being ported to the torch stable ABI. The current main docs still list functional.forced_align. Agent action: pin torchaudio>=2.10 (matched to torch) rather than migrating, unless you need batching or low memory. Failure modes: an environment stuck on 2.9 may lack the op (Inference: the sources do not say whether 2.9 shipped without it; not verified). Limitations: torchaudio is in maintenance, no new features. Version/date: issue edit 2026-01-22; docs fetched 2026-10-24. Limitations: Tutorial rendered with a pre-release torchaudio; availability of the replacement API not checked. Corrected by C3-B866-01 (2026-10-24); where the correction disagrees, the correction governs. Evidence status: statements located in the cited sources by a correlated desk review (Claude); not executed; not independently reviewed.

{
  "citations": [
    {
      "url": "https://docs.pytorch.org/audio/stable/tutorials/ctc_forced_alignment_api_tutorial.html",
      "note": "page warning; intro; Frame-level alignments; Word-level alignments; Technical Details / setup; researcher reasoning"
    },
    {
      "url": "https://raw.githubusercontent.com/m-bain/whisperX/main/README.md",
      "note": "page warning; intro; Frame-level alignments; Word-level alignments; Technical Details / setup; researcher reasoning"
    }
  ]
}