Compare the models
Play the same prompt from OmniVoice and Chatterbox back to back. The page includes scripts, listening cues, and note fields stored in your browser.
Open A/B comparisonLocal voice cloning experiment
Two local voice-cloning models
Serveurperso/OmniVoice-GGUF Q8_0 (base model k2-fsa/OmniVoice) compared with ResembleAI/chatterbox Multilingual V3, using the same eight scripts and the same reference recording, now included below.
Original voice-cloning input
The exact 10.344-second, 24 kHz mono PCM16 reference used for both models. This is the author's real voice recording and is public in this repository.
Download reference WAVListening verdict: OmniVoice was the clear favorite for the author. Chatterbox was judged “not good” overall, with OmniVoice doing better in several respects. Per-clip ratings and the specific dimensions are still open.
8matched scripts
2local model setups
4 GBRTX 2050
$0paid synthesis
Serveurperso/OmniVoice-GGUF · Q8_0
Base model k2-fsa/OmniVoice. Stress tests, per-case GPU telemetry, prompts, and outputs.
Open OmniVoice worksheetResembleAI/chatterbox · Multilingual V3
Per-case timing and memory, model revisions, outputs, and the FP32-to-FP16 constraint on 4 GB VRAM.
Read the Chatterbox report