github machinewrapped/llm-subtrans v1.7.0
Built-in audio transcription (cloud/local)

4 hours ago

Finally crossed the longest-standing TODO off the list - built-in audio transcription, to provide a full end-to-end subtitling flow.

The new Transcribe button opens a dialog where you can select a video file or audio file, which will be analyzed to divide it into silence-bounded chunks and dispatched to the selected transcription service, producing a set of subtitles in the same language as the dialogue that are automatically opened as a translation project.

Transcription Dialog (options)

Note that transcription requires a separate ffmpeg installation.

When the model supports it, speaker identification can be enabled (diarization). This identifies different speakers and tags each line with a speaker index, which is then provided to the translator model to help it understand who is speaking. This helps to avoid some common errors where the model has to guess who is speaking.

Transcription Dialog (running)

With a good model this provides a step-change in accuracy and quality from Whisper-based workflows, and is quite likely to produce more accurate results than text-to-text translation.

Transcribed and translated subtitles

As a bonus, speaker tags in .ass/.vtt files will now be used for translation as well.

Supported transcription services:

  • OpenRouter - provides several good options, with MAI Transcribe-2 being the current standout for speed, price and accuracy
  • Gemini Transcribe - very good and fast, but with tight rate limits and daily caps (large chunks are a must
  • Qwen-ASR - provides on-device transcription, with guided installation of a hardware-appropriate Torch
  • Muse Transcribe - rather slow and does not provide word-level timings, so not the best option here
  • OpenAI - somewhat outdated and largely untested, but feel free to try it

This was a massive feature and transcription should be considered as beta functionality at the moment. In particular, Qwen Local has only been tested with NVidia hardware on Windows... Mac/Linux should work too, but consider yourself a beta tester if you use it.

What's Changed

  • Extend short subtitles on save and preserve subtitle start times by @williamgateszhao in #432
  • Media transcription to timed subtitles (Qwen, OpenRouter, OpenAI, Gemini, Muse) by @machinewrapped in #433
  • Add portable and configurable application paths by @machinewrapped in #430
  • New icon set from GPT-6 Astra

New Contributors

Full Changelog: v1.6.1...v1.7.0

Don't miss a new llm-subtrans release

NewReleases is sending notifications on new releases.