v3.8.2

What's Changed

• feat: expose avg_logprob per segment from ctranslate2 beam search by @Barabazs in #1350
• fix: revert #986 wildcard alignment that broke word-level timestamps (#1220) by @Barabazs in #1367

Full Changelog: v3.8.1...v3.8.2

v3.7.8

Backport of word-level timestamp fixes from v3.8.2.

Bug Fixes

• Restore original CTC forced-alignment (f2609a6): PR #986 caused all words to anchor to the start of the segment window (silence) instead of actual speech. Reverts get_trellis/backtrack to the original PyTorch tutorial implementation. Fixes #1220.
• Fix blank_id hardcoded to 0 (636f298): Broke alignment for HuggingFace models where blank is [pad], not index 0.

Full Changelog: v3.7.7...v3.7.8

v3.6.1

Backport of word-level timestamp fixes from v3.8.2.

Bug Fixes

• Restore original CTC forced-alignment (f2609a6): PR #986 caused all words to anchor to the start of the segment window (silence) instead of actual speech. Reverts get_trellis/backtrack to the original PyTorch tutorial implementation. Fixes #1220.
• Fix blank_id hardcoded to 0 (636f298): Broke alignment for HuggingFace models where blank is [pad], not index 0.

Full Changelog: v3.6.0...v3.6.1

v3.5.1

Backport of word-level timestamp fixes from v3.8.2.

Bug Fixes

• Restore original CTC forced-alignment (f2609a6): PR #986 caused all words to anchor to the start of the segment window (silence) instead of actual speech. Reverts get_trellis/backtrack to the original PyTorch tutorial implementation. Fixes #1220.
• Fix blank_id hardcoded to 0 (636f298): Broke alignment for HuggingFace models where blank is [pad], not index 0.

Full Changelog: v3.5.0...v3.5.1

v3.4.4

Backport of word-level timestamp fixes from v3.8.2.

Bug Fixes

• Restore original CTC forced-alignment (f2609a6): PR #986 caused all words to anchor to the start of the segment window (silence) instead of actual speech. Reverts get_trellis/backtrack to the original PyTorch tutorial implementation. Fixes #1220.
• Fix blank_id hardcoded to 0 (636f298): Broke alignment for HuggingFace models where blank is [pad], not index 0.

Full Changelog: v3.4.3...v3.4.4

v3.3.5

Backport of word-level timestamp fixes from v3.8.2.

Bug Fixes

• Restore original CTC forced-alignment (f2609a6): PR #986 caused all words to anchor to the start of the segment window (silence) instead of actual speech. Reverts get_trellis/backtrack to the original PyTorch tutorial implementation. Fixes #1220.
• Fix blank_id hardcoded to 0 (636f298): Broke alignment for HuggingFace models where blank is [pad], not index 0.

Full Changelog: v3.3.4...v3.3.5

v3.8.1

What's Changed

• Fix: Respect --model_dir and --model_cache_only during alignment by @MrPrayer in #1285
• feat: forward --hf_token to WhisperModel for gated/private model support by @Barabazs in #1351

New Contributors

• @MrPrayer made their first contribution in #1285

Full Changelog: v3.8.0...v3.8.1

v3.8.0

What's Changed

• feat: migrate to pyannote-audio v4 with speaker-diarization-community-1 by @Barabazs in #1349

Special thanks to @borgoat for taking the lead.

Full Changelog: v3.7.7...v3.8.0

v3.7.7

What's Changed

• Optimize assign_word_speakers with interval tree for 228x speedup by @Mr-Neutr0n in #1338
• fix: pass no_repeat_ngram_size and repetition_penalty to CTranslate2 generate() by @RickSanchez93 in #1340
• chore: update type hints by @1carlito in #1342 and #1343
• fix: derive SRT/VTT cue times from word-level timestamps by @Barabazs in #1347

New Contributors

• @Mr-Neutr0n made their first contribution in #1338
• @RickSanchez93 made their first contribution in #1340
• @1carlito made their first contribution in #1342

Full Changelog: v3.7.6...v3.7.7

v3.7.6

What's Changed

• chore: drop python 3.9 support by @Barabazs in #1328

Full Changelog: v3.7.5...v3.7.6