pypi / whisper-s2t-reborn
whisper-s2t-reborn
What changed in whisper-s2t-reborn 1, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-24.
6 breaking3 new APIs and options1 behavior and defaults13 fixes and other
See which of these apply to your code, ranked, in your agent:
npx @unstale/cli initwhisper-s2t-reborn 123 changes
Breaking changes 6
The library now requires huggingface-hub 0.23 or newer.
Five unused decoding settings have been dropped.
This option, which never did anything, has been dropped.
Start times in JSON and TSV output are written as floats, such as 29.0 rather than 29.
Without VAD, a final window under 1 s is transcribed with the preceding window instead of as a separate trailing line.
CTranslate2 4.6.3 is now the minimum supported version, so every install uses its cuDNN-free pure-CUDA convolution path.
New APIs and options 3
The needed aligner model can be predetermined so only that model is downloaded, and it can be run from a full offline folder.
Models can now be loaded fully offline from a folder, enabling use in Docling's offline mode.
The parameter can now be adjusted instead of always using its default.
Behavior and default changes 1
Downloading a model no longer loads PyTorch.
Fixes and other changes 13
Prevents duplicate words from appearing across VAD segments.
Prevents empty transcripts from producing hallucinated text.
Downloads work with every precision setting, can be forced to re-download, and fail immediately when a model is unavailable offline.
The library no longer fails to import in read-only environments.
Short audio waves are padded during extraction and processing.
Abnormally short transcriptions are retried.
Without VAD, fractional max_speech_len values no longer make windows overlap and repeat words.
With word timestamps enabled, very short audio windows no longer cause transcribe_with_vad() to abort with an IndexError.
With word timestamps enabled, very short audio windows no longer cause transcribe() to abort with an IndexError.
Files without an initial prompt are no longer contaminated by prompts from other files in the same batch, making their output identical to transcribing them alone.
Benchmark scripts now exit promptly with a clear error when the worker process dies, rather than hanging indefinitely.
8-, 24-, and 32-bit WAV files are decoded correctly instead of being misread as 16-bit and producing plausible-looking garbage with wrong timestamps.
Word-level timestamps no longer crash on large-v3, large-v3-turbo, and distil-large models; the word aligner falls back to the main model when the default tiny aligner cannot accept its audio features.