# ๐งช PROCESS-2
### Reproducible Speech Analysis & Baseline Modelling Pipeline
[]()
[]()
[]()
[]()
[](https://huggingface.co/datasets/CognoSpeak/PROCESS-2)
[]()
[]()
[](https://doi.org/10.5281/zenodo.19900224)
**Official analysis and experiment pipeline for the PROCESS-2 speech dataset**
---
## ๐ฌ Overview
This repository contains the complete **reproducible analysis pipeline** for the **PROCESS-2 speech dataset**, publicly available on Hugging Face:
๐ https://huggingface.co/datasets/CognoSpeak/PROCESS-2
The pipeline reproduces:
- dataset preprocessing
- statistical analyses
- speech feature extraction
- automatic speech recognition (ASR)
- baseline machine learning models
- large language model (LLM) experiments
All experiments reported in the accompanying publication can be reproduced directly from this repository.
---
## ๐งฌ Research Context
PROCESS-2 supports research in:
- speech-based cognitive assessment
- dementia and MCI biomarkers
- clinical speech analytics
- multimodal AI evaluation
- reproducible machine learning pipelines
The repository follows **open-science and reproducibility best practices**, enabling transparent benchmarking and future comparative studies.
---
## ๐ Repository Structure
```
PROCESS-2/
โ
โโโ codes/ # Analysis & modelling scripts
โ โโโ PROCESS2_gen_audio_info.py
โ โโโ PROCESS2_data_analysis.py
โ โโโ PROCESS2_embed.py
โ โโโ PROCESS2_gen_ASR.py
โ โโโ PROCESS2_BASELINE_class.py
โ โโโ PROCESS2_BASELINE_LLM.py
โ
โโโ envs/ # Conda environments
โ โโโ aconda.yml
โ โโโ embed.yml
โ โโโ torch-gpu.yml
โ
โโโ results/
โ โโโ logs/ # Automatically generated logs
โ
โโโ README.md
```
---
## โ๏ธ Installation
We strongly recommend **Conda** for full reproducibility.
### Create environments
```bash
conda env create -f envs/.yml
```
### Available Environments
| Environment | Purpose |
|-------------|---------|
| **ACONDA** | Preprocessing, statistics, classical ML |
| **embed** | Speech embedding extraction |
| **torch-gpu** | ASR generation & LLM experiments |
Activate before running each stage:
```bash
conda activate
```
---
## ๐ Reproducible Workflow
Run scripts **in order** to reproduce all results.
All commands assume execution from:
```
codes/
```
---
### Step 1 โ Generate Audio Metadata
Computes recording statistics:
- duration
- signal-to-noise ratio (SNR)
- file size summaries
```bash
python PROCESS2_gen_audio_info.py
```
Environment: **ACONDA**
---
### Step 2 โ Statistical Dataset Analysis
Reproduces descriptive statistics reported in the paper.
```bash
python PROCESS2_data_analysis.py
```
Environment: **ACONDA**
---
### Step 3 โ Speech Embedding Extraction
Generates acoustic and linguistic representations for modelling.
```bash
python PROCESS2_embed.py
```
Environment: **embed**
---
### Step 4 โ Automatic Speech Recognition (ASR)
Produces automatic transcripts.
```bash
python PROCESS2_gen_ASR.py
```
Environment: **torch-gpu**
GPU recommended but optional depending on configuration.
---
### Step 5 โ Classical Baseline Models
```bash
python PROCESS2_BASELINE_class.py 12 |& tee -a ../results/logs/PROCESS2_BASELINE_class.txt
```
`12` = number of CPU cores.
Environment: **ACONDA**
---
### Step 6 โ LLM Baseline Experiments
```bash
python PROCESS2_BASELINE_LLM.py 0,1,2,3 |& tee -a ../results/logs/PROCESS2_BASELINE_LLM.txt
```
`0,1,2,3` correspond to GPU device IDs.
Environment: **torch-gpu**
---
## ๐ Outputs
Running the full pipeline produces:
- audio metadata summaries
- dataset statistical analyses
- speech embeddings
- ASR transcripts
- baseline model predictions
- experiment evaluation outputs
- execution logs
Logs are automatically stored in:
```
results/logs/
```
---
## ๐ Reproducibility Statement
This repository provides **full computational reproducibility** for the PROCESS-2 dataset publication.
The workflow ensures:
- deterministic preprocessing
- version-controlled environments
- explicit experiment ordering
- logged execution traces
Researchers may reuse or extend the pipeline for:
- benchmarking studies
- model comparison
- replication experiments
- clinical speech AI research
---
## ๐ค Dataset Access
The PROCESS-2 dataset is hosted on Hugging Face:
๐ https://huggingface.co/datasets/CognoSpeak/PROCESS-2
Please follow dataset licensing and ethical usage guidelines.
---
## ๐ Citation
If you use **PROCESS-2**, please cite **the dataset paper, and code repository**.
---
### ๐ Dataset Paper
Pahar, M., Illingworth, C. H., Mirheidari, B., Elghazaly, H., Peters, F., Young, S., Leung, W. Z., Kaur, L., Blackburn, D., & Christensen, H. (2026). **PROCESS-2: A Benchmark Speech Corpus for Early Cognitive Impairment Detection.** *arXiv preprint arXiv:2605.14888.*
#### BibTeX
```bibtex
@article{pahar2026PROCESS2,
title={{PROCESS}-2: A Benchmark Speech Corpus for Early Cognitive Impairment Detection},
author={Pahar, Madhurananda and Illingworth, Caitlin H. and Mirheidari, Bahman and Elghazaly, Hend and Peters, Fritz and Young, Sophie and Leung, Wing-Zin and Kaur, Labhpreet and Blackburn, Daniel and Christensen, Heidi},
journal={arXiv preprint arXiv:2605.14888},
year={2026}
}
```
### ๐ Code repository
Pahar, M. (2026) *Madhurananda/PROCESS-2: PROCESS-2*. Zenodo. doi:10.5281/zenodo.19900225.
#### BibTeX
```bibtex
@software{pahar_PROCESS2_codes,
author = {Pahar, Madhurananda},
title = {Madhurananda/{PROCESS}-2: {PROCESS}-2},
month = Apr,
year = 2026,
publisher = {Zenodo},
version = {v1.0.0},
doi = {10.5281/zenodo.19900225},
url = {https://doi.org/10.5281/zenodo.19900225}
}
```
---
## ๐ Open Science & Ethics
This repository supports transparent and ethical AI research.
Included:
- โ
code
- โ
experiment pipeline
- โ
reproducibility infrastructure
Not included:
- โ private participant data
All dataset usage must comply with ethical approval and data governance policies.
---
**Reproducible Speech AI Research โข Open Science โข Clinical Machine Learning**