You can deploy this model and start earning money today!

Anirudh45/
flowclean

Name: flowclean
Availability: InStock

Views

No views yet

FlowClean: Conditional Flow Matching for Speech Denoising

Setup

bash

# Create conda env
conda create -n flowclean python=3.11 -y
conda activate flowclean

# Install PyTorch (CUDA 12.1)
pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu121

# Install dependencies
pip install "datasets<4.0" librosa soundfile pyyaml pesq pystoi

# (Optional) Install wandb for experiment tracking
pip install wandb

HuggingFace Token

The dataset is loaded from HuggingFace (JacobLinCool/VoiceBank-DEMAND-16k). If you hit rate limits or access issues:

bash

# Login once
huggingface-cli login
# OR set the env var
export HF_TOKEN=your_token_here
export WANDB_API_KEY= your_token_here

Train

bash

# Single GPU
python train.py --config configs/default.yaml

# Multi-GPU (DDP)
torchrun --nproc_per_node=3 train.py --config configs/default.yaml

DDP is auto-detected via torchrun — no config changes needed. Checkpoints are saved to ./checkpoints/.

With Wandb

In configs/default.yaml, set:

yaml

wandb:
  use_wandb: true
  wandb_token: "your_api_key"   # or set WANDB_API_KEY env var

Then run training as usual. Logs train/loss, train/loss_fm, train/loss_mr, train/lr per step, and epoch/avg_loss per epoch. Inference logs eval/pesq and eval/stoi.

Inference

bash

python inference.py --checkpoint checkpoints/flowclean_best.pt \
                    --ode_steps 10 \
                    --solver euler \
                    --output_dir ./enhanced

Add --eval_metrics to compute PESQ and STOI on the test set.

Project Structure

flowclean/
  config.py          # FlowCleanConfig dataclass
  models/unet.py     # Conditional U-Net backbone
  data/              # VoiceBank-DEMAND HF dataset loader
  losses/stft_loss.py# Multi-resolution STFT loss
  utils/stft.py      # STFT / iSTFT utilities
configs/default.yaml # All hyperparameters
train.py             # Training script
inference.py         # Inference + evaluation

Anirudh45/flowclean

FlowClean: Conditional Flow Matching for Speech Denoising

Setup

HuggingFace Token

Train

With Wandb

Inference

Project Structure

Anirudh45/
flowclean