Transcription via huggingface

Heavily borrows from this tutorial

Using the GUI

First, make sure you have python >= 3.10 and < 3.14 (>= 3.14 will break because of some deprecated numba calls) installed (you can check this by typing python or python3 into your terminal.

Also make sure you have ffmpeg installed on your computer if you will be convering .mp4 files to .mp3 files.

Then, clone this code from GitHub and navigate to the directory where the code lives using your terminal navigating directories for Mac/Linux navigating directories w/ command prompt

Finally, unless you know you have an NVIDIA GPU with the proper CUDA drivers installed, run python3 -m pip install requirements.txt in your terminal.

To run the program, run python3 gui.py and follow the prompts.

To run the code:

Make sure you're running this in an environment that has python >= 3.10 and has the packages in requirements.txt installed.

If you are running this on a machine with a GPU, please install the packages in gpu_requirements.txt instead.

Either way, you will install this using python3 -m pip install -r requirements.txt (or gpu_requirements.txt).

If you're running this on the CoMM Lab machine, you can use the virtual environment that I made a long time ago with much suffering.

Activate it using source /home/asmithh/gbpn/bin/activate.

PLEASE DO NOT EDIT THIS VIRTUAL ENVIRONMENT ON YOUR OWN

Output File Types

You can choose from 3 kinds of output files: VTT (compatible with MAXQDA) TXT (regular plain text file) TSV (tab-separated values file with beginning timestamp for utterance, ending timestamp for utterance, and the utterance. most compatible with Excel & friends but a little hard to work with)

To run the command line script: python3 transcribe_interview.py -i $INFILE -o $OUTFILE -c $CONFIG_FILE

INFILE should be a .mp3 file with your audio you'd like to transcribe.

OUTFILE should be a .tsv, .txt, or .vtt file that has a name you'll successfully associate with your input audio (test.tsv may not be ideal, for example)

CONFIG_FILE is optional; if you don't include a config file it will default to the one that's included with this code.

The config file:

Currently specifies which device to use (0 for the main GPU on the CoMM Lab machine; cpu for CPU (which I do not recommend)) and which transcription model to use (defaults to distil-whisper/distil-medium.en). You can change these settings, but please do so cautiously!!

Useful tips:

To transform m4a to mp3, you can use the ffmpeg command line utility: ffmpeg -i "infile.m4a" -c:v copy -c:a libmp3lame -q:a 4 outf.mp3

This is intended to run on GPU; running the model on CPU is much slower (~20 minutes for ~30 minutes of audio on my 2021 M1 Macbook with 16 GB of RAM).

The whisper class of transcription models may hallucinate when given audio without speech; please keep an eye out for this!

You may also see utterances or sentences duplicated in the transcribed text.

Running on Explorer

Setup

1. Get a Explorer account using these instructions.
2. Go to https://ood.explorer.neu.edu.
3. Click on the "Cluster" tab and then click "Explorer Shell Access".
4. Once you have a terminal window open, request a session: srun --partition=gpu --nodes=1 --pty --gres=gpu:1 --ntasks=1 --mem=4GB --time=02:00:00 /bin/bash
5. Next, clone this repository and enter its directory (cd transcription). To clone a repository, you can find instructions here or you can click the green button at the top of this page and choose the HTTPS tab. Then click the "copy" icon.
6. Now we'll load CUDA; CUDA is a piece of software that lets Python talk to the GPU: module load cuda/12.1
7. Load conda (Anaconda, a Python package manager) using this command: module load anaconda3/2024.06
8. Create a blank Conda environment: conda create -y -n NAME python==3.12.1 where NAME should be something you'll remember is associated with this project.
9. Activate your Conda environment: source activate NAME
10. Install the necessary packages: python3 -m pip install -r gpu_requirements.txt (this will take forever).
11. Back at https://ood.explorer.neu.edu, go to the "Files" tab and open up your home directory. Click the "Upload" button and upload a .mp3 file. Keep track of its name.
12. The fun part: go back to your terminal window and run python3 transcribe_interview.py -i ~/INPUT_FILE_NAME -o ~/OUT_FILE_NAME, where INPUT_FILE_NAME is the name of your input file, and OUT_FILE_NAME is some name for your output file that you'll remember is associated with that input file.

Transcribing once you're set up:

1. Go to https://ood.explorer.neu.edu.
2. Click on the "Cluster" tab and then click "Explorer Shell Access".
3. Once you have a terminal window open, request a session: srun --partition=gpu --nodes=1 --pty --gres=gpu:1 --ntasks=1 --mem=4GB --time=02:00:00 /bin/bash
4. Now we'll load CUDA; CUDA is a piece of software that lets Python talk to the GPU: module load cuda/12.1
5. Load conda (Anaconda, a Python package manager) using this command: module load anaconda3/2024.06
6. Activate your Conda environment: source activate NAME
7. Back at https://ood.explorer.neu.edu, go to the "Files" tab and open up your home directory. Click the "Upload" button and upload a .mp3 file. Keep track of its name.
8. The fun part: go back to your terminal window and run python3 transcribe_interview.py -i ~/INPUT_FILE_NAME -o ~/OUT_FILE_NAME, where INPUT_FILE_NAME is the name of your input file, and OUT_FILE_NAME is some name for your output file that you'll remember is associated with that input file.