liammatt5/GLAM_Web_App

GLAM Pipeline

This repository includes a respiratory audio separation pipeline built around a local SepFormer model and a clinical reasoning/GNN inference path.

Key Files

model.py

Loads the SepFormer separation model using local pretrained files in pretrained_sepformer.

Wraps encoder/masknet/decoder into UnifiedSepFormer.

interface.py

Provides shared audio loading, separation, and saving helpers.

Includes:

• —separate_audio_file(...)
• —predict_sources(...)
• —save_separated_sources(...)

for pipeline reuse.

Still supports the existing Gradio UI workflow via:

• —separate_audio(...)

pipeline.py

Runs separation over a directory of mixture audio files.

Writes separated WAVs and waveform PNGs to an output folder.

Produces:

• —pipeline_summary.json
• —pipeline_summary.csv

gnn.py

Defines the GNN model architecture used for respiratory audio reasoning.

Provides utilities for:

• —loading a GNN checkpoint
• —estimating breathing rate
• —patient state tracking
• —clinical alert generation

reasoning_pipeline.py

Processes GNN run outputs:

• —window_report.json
• —window_report.csv

Generates reasoning summary CSV/JSON files.

Includes visualization helpers for per-patient summary plots.

full_pipeline.py

Runs the complete flow:

1. 1.Separate mixture audio files
2. 2.Run Wav2Vec2 + GNN inference
3. 3.Save per-run artifacts
4. 4.Aggregate reasoning summaries

Running a Simulation

To test the system using the provided sample data:

1. 1.Open the Patient Registration page.
2. 2.Navigate to the Sample_Audio folder.
3. 3.Upload S1, S2, and S3 as individual patient recordings.
4. 4.Click Register to register each patient.

Next:

1. 1.Open the Separation page.
2. 2.Upload Sample0mix.wav from the Sample_Audio folder as the mixed audio recording.
3. 3.Select the registered patients (S1, S2, and S3).
4. 4.Click Run Separation.

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference