avm-voice-weights

Weights downloaded on first run by the avm-voice realtime singing voice changer (RVC decoder + noise-robust front end). The app fetches these files itself; you do not need to download them manually.

file	what it is	origin / license
encoder_v4.pt	Streaming, noise-robust fine-tune of the RVC HuBERT/ContentVec content encoder (12 layers, fp32). Replaces the base encoder weights at load time.	Fine-tuned by this project from the RVC hubert_base (MIT).
voice_mask_v4.pt	Small speaker-conditioned (ECAPA 192-d, FiLM) causal mask that keeps the enrolled user's voice in the front separator. The enrollment vector of the person who trained it has been removed; the app conditions it on each user's own enrollment made locally.	Trained by this project.
ecapa_1s.ts	TorchScript trace of the ECAPA-TDNN speaker encoder, 1 s causal window, used for the user-voice gate and enrollment.	Traced from speechbrain/spkrec-ecapa-voxceleb (Apache-2.0).

No user recordings, voice embeddings or enrollment vectors are included in this repository. HuBERT (hubert_base/*) and RMVPE (rmvpe.pt) are downloaded by the app from the public lj1995/VoiceConversionWebUI repository and are not re-hosted here.