PS4: Proxy-Supervised Joint Training for Real Target Speaker Extraction
Paper โข 2607.08111 โข Published
Rust/Candle-friendly safetensors conversion of the PS4 target speaker extraction checkpoint.
This repository contains only ps4.safetensors, converted from
checkpoint_epoch037.pt with scripts/convert_ps4.py from the
UltimateTranscript repository. Optimizer state was dropped and
num_batches_tracked buffers were removed.
Model summary: BSRNN separator with a joint ECAPA-TDNN speaker encoder, 16 kHz, 27,654,874 parameters. The checkpoint includes the torchaudio mel filterbank and window buffers used by the original inference code.
The original model was trained on REAL-PS4, which aggregates AISHELL-4, AliMeeting, AMI, and CHiME-6. Refer to the original model card for details.