github machinewrapped/llm-subtrans v1.7.2
Improved transcription quality

3 hours ago

See the v1.7.0 release notes for full details about integrated transcription, the major new feature in 1.7.x.

Transcription improvements

This release features significant improvements to the transcription flow, particularly around the way subtitles are divided/grouped into lines. Using a mix of sources, each provider was assessed and analyzed for weaknesses, resulting in some global and specific improvements.

Each provider uses its most reliable signal for content. Sentences that could be dropped or scrambled by the processing are more reliably complete and correct, and timing is validated and adjusted if necessary to avoid over-long or over-short lines.

Most improved is Qwen Local, which is now also fully compatible with macOS.

Multi-lingual transcriptions are handled more intelligently, with language-aware timing validation. More aggressive timing adjustments can be enabled with a new timing_correction_factor (from 0 to 1).

Other improvements to the handling of line breaks, dialog markers and filler word removal should produce translations that need no further work more of the time.

Transcription provider support

  • OpenRouter: MAI Transcribe 2 is current state of the art - fast, accurate, reliable and cheap. This should be your first choice.
  • Qwen Local: Runs entirely on your machine, so effectively free. Coverage and quality is good, but timing info is incomplete so some subtitles may be misaligned.
  • Gemini Transcribe: Fast and reasonably cheap, but prone to silently dropping large chunks of dialogue.
  • OpenAI: Old models, expiring API credits... no good reason to use this unless you have some API credits that need burning.

Transcription-aware instructions

An updated instructions file supersedes the old Whispered instructions, with instructions better tailored to translating subtitles sourced from speech recognition. Selected automatically when opening a translation project from the transcription flow.

Qwen Local transcription is now a fully external installation

qwen_asr is no longer bundled with the package - this reduces the package size greatly, as it pulled in many dependencies that are not necessary if you never use Qwen Local, and avoids issues with packages that are pulled in with less permissive licenses.

The Transcription Settings tab in the Settings dialog will walk you through the install process, which is basically the same as the previous Torch installation with a few more packages to download. If you previously set up an external Torch, just point the setup at its location and the additional packages will be installed there.

Qwen Local transcription on macOS

Torch installation was fixed for macOS, and compatibility issues between qwen-asr and pyside6 were resolved.

Optimised startup time

The initialisation of expensive provider SDK's is now deferred, which drastically reduces start up times. A background task warms the selected default provider once the main window is opened.

Improvements to the settings dialogues prevent freezing when other providers are selected, with background initialization and supported model retrieval.

Miscellaneous fixes and improvements

  • Selection fixes: Split Batch is available again when one line is selected, and Translate Selection covers exactly what's selected.
  • Merged subtitles no longer disappear from the subtitle view, and undoing a merge restores the lines correctly.
  • If the translation provider can't be created , translation stops with an error rather than failing for every scene.
  • Subtitle files with duplicate line numbers are renumbered when loaded.
  • A rejected project file no longer clears the subtitles that are already loaded.
  • Fixed an empty proxy setting for OpenRouter and Requesty.

Main changes

Full Changelog: v1.7.0...v1.7.2

Note that v1.7.1 was superseded without becoming an official release.

Don't miss a new llm-subtrans release

NewReleases is sending notifications on new releases.