Trelis Debuts Chorus-v1-GGML For Local Voice Separation
Trelis recently released a specialized speech transcription model that handles overlapping conversations between two participants. The system processes audio clips locally without relying on external cloud servers.
Built as an optimized adaptation of a popular open speech engine, this project focuses on accurate speaker separation for short recordings. It targets users who need reliable local transcription without complex server setups.
Model Size: from 474MB & VRAM GPU: requirements vary
Local speech separation and quantization
- Processes up to thirty seconds of English audio in a single run.
- Includes multiple file formats ranging from four hundred fifty megabytes to one point six gigabytes.
- Separates overlapping voices using a custom command line switch.
- Runs natively on standard processors through a patched audio pipeline.
- Offers ready-made scripts for both python coding and local web server hosting.
Audio archivists and media professionals managing interview recordings can use these tools to quickly isolate individual speakers and organize tracks. Running these files locally keeps sensitive conversations secure while providing structured data for standard editing workflows.
Technical constraints and format decisions
The project relies on the established GGML file structure because the underlying audio processing library has not transitioned to newer standards.
"Chorus v1 is a LoRA fine-tune of whisper-large-v3-turbo with two extra special tokens,"
noted the developer in a readme.
This design choice requires a minor software patch to recognize custom speaker markers. Mixed precision compression was tested initially, though current loading restrictions prevent those files from initializing correctly. Users must manually split longer recordings and maintain consistent labels between segments.
You can access the quantized model files and installation instructions here to start testing your own clips.
