GigaAM and ComfyUI: video transcription and translation

This Russian tutorial explains a ComfyUI workflow for transcribing Russian speech, correcting terminology, translating text and generating a replacement audio track. GigaAM handles transcription, while custom OreX nodes load long recordings and combine the translated speech with the original video. The speaker also shows how to stop after transcription and keep the text for other uses.

Setup includes installing two node packages and their requirements. The first model download requires a Hugging Face read token; the diarization model also requires accepting its agreement. The demonstrated node supports timestamps and speaker labels. The speaker warns that this workflow's diarization setting produces an error with a single speaker.

LM Studio and Ollama provide alternative routes for text correction and translation. A terminology dictionary guides corrections, and the example splits text into 3,000-token chunks at sentence boundaries for a local LLM. Ollama cloud models and manual editing through Gemini are optional online routes.

The dubbing examples compare CosyVoice and OmniVoice on English and Chinese output, including a two-speaker test. The speaker prefers CosyVoice's pronunciation stability but reports faster synthesis and closer voice cloning with OmniVoice. These are his observations, and he asks fluent listeners to assess pronunciation. Long files have preview and trimming limitations, and uploaded recordings remain in ComfyUI's input folder until manually deleted.

The node repositories are public; the author distributes the full saved workflow and extra instructions through Boosty.