AffCom Lab: Secure Whisper Environment

This is the AffCom Lab Edition of the GPU-accelerated Whisper environment.

It is built on top of the public jmgirard/audio-whisper image but adds a secure "Lab Layer" that allows you to connect directly to the AffCom Datasets and Projects network shares from inside the container.

Features

• GPU Acceleration: Runs OpenAI's Whisper model on the lab machine's NVIDIA GPU.
• Pre-Loaded Models: Uses the Whisper models already stored on the server (no downloads required).
• Safety Rails:
  • /mnt/datasets: Mounted Read-Only (You cannot accidentally delete raw data).
  • /mnt/projects: Mounted Read-Write (Save your transcripts here).

---

Prerequisites

1. Lab Machine: This image is designed to run on AffCom lab workstations.
2. Docker Desktop: Installed and running.
3. NVIDIA Drivers: Ensure the machine's drivers are up to date.

---

Setup & Launch

1. Download the Repository

Open a terminal (PowerShell) and run:

git clone https://github.com/affcomlab/lab-whisper.git
cd lab-whisper

2. Start the Environment

Open your terminal (PowerShell) inside the lab-whisper folder and run:

docker compose run --rm whisper-lab

• The first time you run this, it will take a moment to build the lab image (lab-whisper:latest).
• Subsequent runs will be instant.

3. Connect to Lab Drives

Once R launches, you will see a message: Loaded AffCom Lab Utils. Run the connection function:

connect_lab_drives()

You will be prompted for your KU Online ID and Password.

• Note: Your password is typed securely (it won't show on screen) and is never saved to disk. It exists only in RAM for this session.

If successful, you will see:

✅ Mounted: /mnt/datasets (Read-Only) ✅ Mounted: /mnt/projects (Read-Write)

---

Usage Example

You can analyze audio files directly from the server using the pre-downloaded models.

library(audio.whisper)
library(openac)

# 1. Login (if you haven't already)
connect_lab_drives()

# 2. Load Whisper Model from Server
# We use the models stored in datasets/whisper so you don't have to download them.
# Available: "ggml-base.bin", "ggml-small.bin", "ggml-medium.bin", "ggml-large-v3.bin"
model_path <- "/mnt/datasets/whisper/ggml-base.bin"
model <- whisper(model_path, use_gpu = TRUE)

# 3. Transcribe a file from the PROJECTS drive
# (Replace this path with your actual study folder)
audio_path <- "/mnt/projects/MyStudy/Wave1/subject_001.wav"
transcript <- predict(model, newdata = audio_path)

# 4. Save results back to the server
write.csv(transcript, "/mnt/projects/MyStudy/Wave1/subject_001_transcript.csv")

---

Troubleshooting

"Connection failed" or "Host is down"

• Check Credentials: Did you type your password correctly? (The prompt does not show asterisks ***).
• Check NetID: Use just your username (e.g., jdoe), not your full email.

"Permission Denied" when saving files

• Wrong Folder: Check where you are trying to save.
  • You cannot save to /mnt/datasets (Read-Only).
  • You can save to /mnt/projects (Read-Write).

"Fatal error: you must specify '--save'..."

• You likely ran docker-compose up.
• Always use the run command: docker compose run --rm whisper-lab