How resemble-ai/resemble-enhance works
Resemble Enhance is a speech-improvement application: users provide audio and receive denoised or perceptually enhanced audio. The documented workflow offers a denoise-only option or enhancement, which includes denoising; the web demo is another entry point. Training is a separate capability: prepare foreground/background/RIR data, train the denoiser, then train the enhancer’s autoencoder/vocoder and latent CFM stages. The graph preserves those stages and key model and runtime layers. Some CLI, training, and data wiring is unsampled; those relationships are included only where the README or source index supports them.
Main entry points: enhancer.py Enhancer Model · inference.py Enhance API · enhancer/train.py Enhancer Training · denoiser/train.py Denoiser Training
Components 25
User Interfaces
- CLI Workflow
resemble_enhance/enhancer/__main__.py - Web Demo
app.py
Audio Inference
- Enhance API
resemble_enhance/enhancer/inference.py - Denoise API
resemble_enhance/denoiser/inference.py - Chunked Inference
resemble_enhance/inference.py - Model Download
resemble_enhance/enhancer/download.py
Data and Training
- Audio Dataset
resemble_enhance/data/dataset.py - Audio Distortion
resemble_enhance/data/distorter/base.py - Denoiser Training
resemble_enhance/denoiser/train.py - Enhancer Training
resemble_enhance/enhancer/train.py - Training Loop
resemble_enhance/utils/train_loop.py
Models and Runtime
- Audio Parameters
resemble_enhance/hparams.py - Mel Features
resemble_enhance/melspec.py - Denoiser Model
resemble_enhance/denoiser/denoiser.py - UNet
resemble_enhance/denoiser/unet.py - Enhancer Model
resemble_enhance/enhancer/enhancer.py - IRMAE Autoencoder
resemble_enhance/enhancer/lcfm/irmae.py - Flow Matching
resemble_enhance/enhancer/lcfm/cfm.py - Latent CFM
resemble_enhance/enhancer/lcfm/lcfm.py - UnivNet Vocoder
resemble_enhance/enhancer/univnet/univnet.py - Audio Discriminator
resemble_enhance/enhancer/univnet/discriminator.py - Distributed Engine
resemble_enhance/utils/engine.py - Distributed Runtime
resemble_enhance/utils/distributed.py - Logging
resemble_enhance/utils/logging.py
Other components
- Audio User
Connections 29
- Audio User → CLI Workflow: provides audio
- Audio User → Web Demo: provides audio
- CLI Workflow → Enhance API: runs enhancement
- Web Demo → Enhance API: invokes
app.py - Enhance API → Model Download: loads model
resemble_enhance/enhancer/inference.py - Enhance API → Chunked Inference: runs inference
resemble_enhance/enhancer/inference.py - Denoise API → Chunked Inference: runs inference
resemble_enhance/denoiser/inference.py - Chunked Inference → Audio Parameters: reads settings
resemble_enhance/inference.py - Enhance API → Enhancer Model: loads model class
resemble_enhance/enhancer/inference.py - Denoise API → Denoiser Training: loads model class
resemble_enhance/denoiser/inference.py - Denoiser Training → Audio Dataset: uses data
resemble_enhance/denoiser/train.py - Enhancer Training → Audio Dataset: uses data
resemble_enhance/enhancer/train.py - Audio Dataset → Audio Distortion: uses distorters
resemble_enhance/data/dataset.py - Denoiser Training → Denoiser Model: trains
resemble_enhance/denoiser/train.py - Enhancer Training → Enhancer Model: trains
resemble_enhance/enhancer/train.py - Denoiser Training → Training Loop: uses loop
resemble_enhance/denoiser/train.py - Enhancer Training → Training Loop: uses loop
resemble_enhance/enhancer/train.py - Denoiser Model → Mel Features: extracts features
resemble_enhance/denoiser/denoiser.py - Denoiser Model → UNet: predicts with
resemble_enhance/denoiser/denoiser.py - Enhancer Model → Denoise API: denoises input
resemble_enhance/enhancer/enhancer.py - Enhancer Model → Mel Features: extracts features
resemble_enhance/enhancer/enhancer.py - Enhancer Model → Latent CFM: generates latent
resemble_enhance/enhancer/enhancer.py - Enhancer Model → UnivNet Vocoder: reconstructs audio
resemble_enhance/enhancer/enhancer.py - Latent CFM → IRMAE Autoencoder: uses autoencoder
resemble_enhance/enhancer/lcfm/lcfm.py - Latent CFM → Flow Matching: uses flow model
resemble_enhance/enhancer/lcfm/lcfm.py - Enhancer Training → Audio Discriminator: configures discriminator
resemble_enhance/enhancer/train.py - Training Loop → Distributed Engine: uses engine
resemble_enhance/utils/train_loop.py - Distributed Engine → Distributed Runtime: initializes distributed
resemble_enhance/utils/engine.py - Logging → Distributed Runtime: guards logging
resemble_enhance/utils/logging.py