Quick Start

Training

Run the log-magnitude model (recommended starting point):

python -m libdse.train.simpleAE_logmag_nc

The checkpoint with the best validation loss is saved to models/simple_autoencoder_logmag_spec_noisy_clean.

Monitor training with TensorBoard:

tensorboard --logdir runs

Gradio Demo

A pre-trained checkpoint is included in models/. Launch the demo:

python -m libdse.showcases.simpleAE_logmag_nc

Open http://127.0.0.1:7860 in a browser. Two tabs are available:

• Denoise — upload any audio file; spectrograms of the noisy input and the enhanced output are shown side-by-side.

• Noise mix — upload clean speech, select a DEMAND environment, set a target SNR, and listen to the synthesised noisy mixture.

Running Tests

uv run pytest

Using the API directly

from pathlib import Path
import librosa
import torch
from libdse.nets import VanillaAutoEncoder
from libdse.data.features import LogMagnitudeSpectrumExtractor
from libdse.data.noise import DEMANDNoiseDataset, DEMANDNoiseType
from libdse.data.librispeech import LibriSpeechDataset
from torch.utils.data import DataLoader

# --- Build noise dataset -------------------------------------------------
noise_ds = DEMANDNoiseDataset(
    entry_point=Path("data/noise/DEMAND"),
    noise_types=DEMANDNoiseType.ALL,
)

# --- Build feature extractor ---------------------------------------------
extractor = LogMagnitudeSpectrumExtractor(
    sampling_rate=8_000,
    window_length=256,
    hop_length=128,
    noise=noise_ds,
)

# --- Wrap the LibriSpeech corpus -----------------------------------------
ds = LibriSpeechDataset(
    entry_point=Path("data/train-clean-100"),
    extractor=extractor,
    sample_rate=8_000,
)
loader = DataLoader(ds, batch_size=256)

# --- Define the model ----------------------------------------------------
model = VanillaAutoEncoder(
    input_dim=ds.sample_shape[0],   # 129 for window_length=256
    latent_dim=180,
    hidden_layer_struct=[2048, 500],
)

# --- Single forward pass -------------------------------------------------
noisy_batch, clean_batch = next(iter(loader))
enhanced = model(noisy_batch.float())
print(enhanced.shape)   # (256, 129)