db84llm

Overview

[doc] [sheet]

Dependencies

• Python
• Jupyter Notebook
• openai/whisper
• Whisper JAX (see below)
• yt-dlp
• sox

Update List of Debates

curl -L 'https://docs.google.com/spreadsheets/d/1SgFmh0L-VZMappWinTkA628Le5dgT2hmosngM2jDdlo/export?exportFormat=csv' --output list_of_debates.csv

Audio Files

Compiled a short list of debates to look into: list_of_debates.csv. Will most likely use a Google Spreadsheet API to dynamically get these in the future.

# yt-dlp
yt-dlp -x --audio-format wav "https://www.youtube.com/watch?v=example_video_id"

Transcripts

For our purposes, I found Whisper JAX to suit our needs. I've detailed a step-by-step guide on how to setup Whisper JAX along with a couple template files to test it out.

Setup

Follow along the Whisper JAX README file. I've listed a couple essential steps below:

Step 1: Install JAX
Navigate to the JAX page and look for the instructions.

Hardware	Instructions
CPU	pip install -U jax
NVIDIA GPU	pip install -U "jax[cuda12]"
Google TPU	pip install -U "jax[tpu]" -f https://storage.googleapis.com/jax-releases/libtpu_releases.html
AMD GPU	Use Docker or build from source.
Apple GPU	Follow Apple's instructions.

Apple GPU
I primarily use Anaconda environments instead of Python Virtual Environment (venv). The instructions here choose to use venv. I found this Medium article where it sets up a conda env specifically to build PyTorch on Metal/MPS. Lmk if you want that article.

python3 -m venv ~/jax-metal # optional
source ~/jax-metal/bin/activate # optional
python -m pip install -U pip
python -m pip install numpy wheel
python -m pip install jax-metal

jaxlib Compatibility

pip install -U jaxlib jax
ENABLE_PJRT_COMPATIBILITY=1 python -c 'import jax; print(jax.numpy.arange(10))'

^^ Remember this, we'll need to create an environment variable later.

Step 2: Installing Whisper JAX

pip install git+https://github.com/sanchit-gandhi/whisper-jax.git

Step 3: Running Whisper
Navigate to the whisper_jax.ipynb file. If you haven't already, go ahead and pip install sox. The debate I'm primarily using for testing is Dartmouth RR 2024 - Round 4 - Michigan PD vs Dartmouth BC, and I've uploaded each speech individually already so you don't need to run the second cell where it slices the audio file. You still need to run the first cell though.

After that, in the same directory, create a file called .env if not already created. In this file, put in the environment variable mentioned earlier:

ENABLE_PJRT_COMPATIBILITY=1

This way, we can load the environment variable into our current notebook. Everything else should be able to run straight down. I have a dummy cell using Whisper on audio_trimmed which is just a one minute audio clip from some random speech. You can choose to run it or not. Also, I've run into a couple random errors that I don't quite understand, such as XlaRuntimeError: UNKNOWN. To solve this, I restart my kernel, maybe pip install a package again, and just try until it works.