You can deploy this model and start earning money today!

yskim3271/
lacosenet-voicebank-demand

Name: lacosenet-voicebank-demand
Availability: InStock

Views

No views yet

LaCo-SENet VoiceBank+DEMAND Checkpoints

This repository contains the VoiceBank+DEMAND checkpoints for LaCo-SENet, a latency-configurable streaming speech enhancement model.

Paper: Latency-Configurable Streaming Speech Enhancement via Asymmetric Temporal Padding Code: yskim3271/LaCo-SENet

The release includes the best-validation checkpoints for 9 latency configurations and 3 random seeds, for a total of 27 checkpoints. Each run directory contains:

model_<step>.th: PyTorch model weights.
.hydra/config.yaml: model and training configuration required by the loader.
.hydra/overrides.yaml: Hydra overrides used for the run.

The checkpoints were trained on VoiceBank+DEMAND 16 kHz using JacobLinCool/VoiceBank-DEMAND-16k.

Training Setup

GPU: single NVIDIA GeForce RTX 5090 per run
Steps: 400,000
Batch size: 8
Optimizer: AdamW
Learning rate: 5e-4
Typical wall-clock time: about 29-31 hours per run

Checkpoints

Config	Latency	Seed 2039	Seed 42	Seed 7
M1	12.5 ms	`M1_12.5ms/s2039/model_279000.th`	`M1_12.5ms/s42/model_130000.th`	`M1_12.5ms/s7/model_130000.th`
M2	25.0 ms	`M2_25.0ms/s2039/model_147000.th`	`M2_25.0ms/s42/model_169000.th`	`M2_25.0ms/s7/model_130000.th`
M3	37.5 ms	`M3_37.5ms/s2039/model_99000.th`	`M3_37.5ms/s42/model_252000.th`	`M3_37.5ms/s7/model_130000.th`
M4	50.0 ms	`M4_50.0ms/s2039/model_97000.th`	`M4_50.0ms/s42/model_123000.th`	`M4_50.0ms/s7/model_189000.th`
M5	62.5 ms	`M5_62.5ms/s2039/model_99000.th`	`M5_62.5ms/s42/model_123000.th`	`M5_62.5ms/s7/model_326000.th`
M6	75.0 ms	`M6_75.0ms/s2039/model_163000.th`	`M6_75.0ms/s42/model_260000.th`	`M6_75.0ms/s7/model_261000.th`
M9	100.0 ms	`M9_100.0ms/s2039/model_249000.th`	`M9_100.0ms/s42/model_215000.th`	`M9_100.0ms/s7/model_264000.th`
M13	150.0 ms	`M13_150.0ms/s2039/model_240000.th`	`M13_150.0ms/s42/model_214000.th`	`M13_150.0ms/s7/model_171000.th`
M7	200.0 ms	`M7_200.0ms/s2039/model_184000.th`	`M7_200.0ms/s42/model_159000.th`	`M7_200.0ms/s7/model_315000.th`

See manifest.json for SHA256 checksums and per-checkpoint test metrics.

Usage

Install the code repository and dependencies first:

bash

git clone https://github.com/yskim3271/LaCo-SENet.git
cd LaCo-SENet
pip install -r requirements.txt

Then download this checkpoint repository and pass a run directory plus checkpoint filename to the evaluation or enhancement scripts:

bash

python -m src.evaluate --chkpt_dir /path/to/checkpoints/M6_75.0ms/s2039 \
  --chkpt_file model_163000.th \
  --device cuda

For streaming RTF measurement:

bash

python -m src.measure_rtf --chkpt_dir /path/to/checkpoints/M6_75.0ms/s2039 \
  --chkpt_file model_163000.th \
  --chunk_size 8 \
  --use_onnx

Citation

bibtex

@inproceedings{kim2026lacosenet,
  title     = {Latency-Configurable Streaming Speech Enhancement via Asymmetric Temporal Padding},
  author    = {Kim, Yunsik and Chung, Yoonyoung},
  booktitle = {Proc. Interspeech},
  year      = {2026}
}

yskim3271/lacosenet-voicebank-demand

LaCo-SENet VoiceBank+DEMAND Checkpoints

Contents

Training Setup

Checkpoints

Usage

Citation

yskim3271/
lacosenet-voicebank-demand