Skip to content

CHANGELOG.md

Latest commit

 

History

History
83 lines (66 loc) · 7.24 KB

CHANGELOG.md

File metadata and controls

83 lines (66 loc) · 7.24 KB
Raw

Changelog

v1.2.7 — 2026-08-24

Added

  • NLLB-200 offline translation engine on CTranslate2 (translation_engine: "nllb" / --translation-engine nllb): fully offline, any mapped language pair, GPU-accelerated (int8_float16 on CUDA, int8 on CPU). The model (~650 MB) downloads once, automatically, at the start of the first run that needs it, into the models/ directory. Zero new dependencies — CTranslate2 runs it, tokenizers (already present via faster-whisper) reads its tokenizer, huggingface_hub (likewise) downloads it. Note: the NLLB model weights are CC-BY-NC-4.0 (non-commercial use only) — the engine logs this at every load; see README, "Offline translation (NLLB)".
  • translation_fallback config field / --translation-fallback flag: what happens when a Google GTX batch fails outright — nllb (translate offline; default), mymemory (the old behaviour), or none (keep the source text).
  • translation_model config field: which NLLB conversion to use — a HuggingFace repo ID, a folder name under models/, or a full path.
  • --self-check now reports whether the NLLB model is on disk.
  • max_chunk_s / min_chunk_s config fields (--max-chunk-s / --min-chunk-s): the chunked-inference sizes, previously hardcoded, are now tunable. Validated before any audio work (0 < min < max ≤ 30s).
  • Add button in the SRT Lines toolbar: creates the first line of a from-scratch subtitle file, prefilled from the playhead. Deliberately enabled only while the list is empty — once any line exists, Split's free-form times already place a new line anywhere (including gaps), and a second insertion affordance would only duplicate it.

Changed

  • Google GTX rate-limit handling: HTTP 429/503 responses back off on a longer ladder (2 s, 8 s) and honour Retry-After (capped at 30 s) instead of retrying at 0.25 s — retrying a rate limiter that fast only deepens the hole the IP is in. Batch requests are additionally paced 0.4 s apart so a 3,000-cue file no longer presents the burst signature that provokes throttling.
  • Translation failure logging no longer floods: the first failed batch logs a WARNING with its cause, subsequent failures log at DEBUG, and the run ends with one summary WARNING carrying the totals and the likely diagnosis. Previously a throttled IP produced one WARNING per batch — ~80 for a typical long recording.
  • If NLLB is configured as the fallback but its model cannot be fetched (offline machine), the run warns once and continues with translation_fallback: "none" rather than failing. NLLB as the primary engine still fails loudly when unavailable — you asked for it by name.

Fixed

  • Chunked inference no longer discards short utterances. Speech regions briefer than min_chunk_s (2s) were silently dropped — deleting every short exclamation from the transcript, measured at ~4 minutes of speech lost in a 10-minute sparse-dialogue excerpt. Such regions are now transcribed whole (short_region in the chunk-plan log, which reports how many were kept). min_chunk_s now governs where cuts may land, not which speech exists.
  • A translation batch that failed after Google's own retries always fell back to MyMemory, whose output quality is not usable for subtitles and whose per-cue round-trips added ~50 s per failed batch. Failure handling is now configurable and defaults to an engine that produces usable output offline.

v1.2.6 — 2026-08-14

Added

  • models\ directory beside the executable for locally-converted CTranslate2 models; a bare folder name resolves against it.
  • Conversion guidance now prints a ready-to-run ct2-transformers-converter command with the real output path filled in.
  • Error messages are selectable and have a copy button.
  • Patch distribution for CUDA installs (Create-gensrt-patch.ps1, tools/make_patch.py), so an update is ~13 MB rather than a full re-download.

Changed

  • Model validation reports HuggingFace 401 responses accurately: private, gated and nonexistent repositories are indistinguishable from outside, and GenSRT no longer claims otherwise.
  • Certificate-verification failures explain the likely cause and what to try.
  • A misconfigured REQUESTS_CA_BUNDLE / CURL_CA_BUNDLE / SSL_CERT_FILE is named explicitly, with the offending values.
  • gensrt-config.json and gensrt-known-models.json are created beside the executable rather than in the working directory. Existing files are not moved.

Fixed

  • Text selection was disabled at the pywebview window level (text_select=False), below anything CSS could override.
  • Local models were only searched for beside the executable, so a model in a project folder was invisible when running from source.
  • Multi-line messages collapsed into a single paragraph.

v1.2.5 — 2026-08-12

Added

  • CPU-only installer variant (Pack-gensrt.ps1 -Variant cpu), alongside the CUDA build.
  • --target-language CLI flag. Translation to non-English targets now works.
  • --self-check verifies an installation is complete: imports every module, runs the bundled FFmpeg, loads the CUDA libraries by name, and tests HTTPS to HuggingFace. The build refuses to produce an installer that fails it.
  • --dump-segments and --debug-chunks export per-cue decoder telemetry and per-chunk audio.
  • Cue numbers in the cue list.
  • WebVTT output alongside SRT.

Changed

  • PyTorch removed. GPU detection uses CTranslate2 instead. Substantially smaller downloads for every user.
  • Offline translation engines (NLLB-200, MarianMT) removed. Both could only produce English and required PyTorch. Old configs naming them get an explanation.
  • Custom-model validation now checks for CTranslate2 format, not just that the repo exists.
  • Model validation uses the same TLS trust store as model downloads.
  • Default subtitle line length is 42 characters (was 84).
  • device defaults to auto and is honoured when set explicitly.
  • Recommended Malayalam model is now adalat-ai/ct2-whisper-medium-ml-rmft.

Fixed

  • Subtitle text past two lines was silently discarded.
  • Cues could overlap each other after the minimum-duration floor was applied.
  • GUI footer was clipped at 125%+ display scaling (#1, reported by @moob158).
  • "device": "cpu" was ignored — the GPU probe overwrote it unconditionally.
  • Burn-in failed on filenames containing [, ], ,, ; or ', silently.
  • Dropping a language-variant SRT (clip.ml.srt) loaded clip.srt instead.
  • Google translation batching broke on non-English targets: the batch delimiter was itself being translated, leaving ~93% of cues untranslated.
  • pyproject.toml declared MIT; GenSRT is AGPL-3.0.
  • python -m gensrt ran the CLI on import and always exited 0.
  • Chunk diagnostics were written to a directory named after the temp audio file.

Known limitations

  • Fine-tuned models emit very short spurious cues at chunk boundaries (median 60 ms). Cleanup is candidate work for v1.3.
  • A � at the end of a line means the model stopped generating mid-character. The text before it is valid.
  • Monolingual fine-tunes render other-language passages phonetically rather than skipping them.
  • Cue boundaries within a chunk are Whisper's own and occasionally split mid-word.

v1.2.1 and earlier

See the release history.