RecorrNet-v2 — AISHELL-1 Mix ver2 + AliMeeting
Single-channel RecorrNet-v2 (direct 9-tap filter head) continued on AISHELL-1 Mix ver2 (clean + both, 1–5 speakers) plus ~10% AliMeeting Train (far CH0 mix, headset stems).
- Architecture: RecorrNet-v2 (
1ch_aishell1mix_ver2_alimeeting_recorr_v2) - Init: RecorrNet-v2 AISHELL continue (not WSJ-init)
- Checkpoint: exported
best_model.pth(epoch 9, step 20000) - Code: vu-duy-tung/byd-speech-separation (
recorrnetbranch)
This replaces the earlier AISHELL-1 Mix ver1 epoch 7 RecorrNet weights as the current RecorrNet checkpoint.
Previous ver1 epoch-7 weights: playwithmino/recorrnet-v2-1ch-aishell1mix-ver1-epoch7.
Load
from sr_corrnet import SSInference
model = SSInference.from_pretrained(
checkpoint_path="playwithmino/recorrnet-v2-1ch-aishell1mix-ver2-alimeeting",
device="cuda:0",
)
result = model.process_file("mix.wav", output_dir="output/")
Use this repo’s config.yaml with these weights. Do not mix with the SR-CorrNet or ver1-epoch7 configs.
Files: model.pt (weights only), config.yaml, metadata.json.