Quick script to pull down a youtube video and dump a whispertranscription (requires whisper, yt-dlp, ffmpeg)

yt-translate.sh

#!/usr/bin/bash

MODEL_PATH="/models/whisper/ggml-large.bin"
#MODEL_PATH="/models/whisper/ggml-base.bin"

YT_ID=$1

# Check if the user has provided a YouTube video ID

if [ -z "$YT_ID" ]; then
    echo "Please provide a YouTube video ID"
    exit 1
fi

echo "- Downloading video $YT_ID"

# check if mp3 already exists and download it if not

if [ ! -f /tmp/$YT_ID.mp3 ]; then
    yt-dlp $YT_ID -x --audio-format mp3 -o /tmp/$YT_ID.mp3
fi

# convert to whisper-friendly wav format
ffmpeg -y -i "/tmp/$YT_ID.mp3" -acodec pcm_s16le -ac 1 -ar 16000 "/tmp/$YT_ID.wav"

# transcribe

whisper.bin -f "/tmp/$YT_ID.wav" -m $MODEL_PATH -pp -t 24 -pc -l en -otxt -of $YT_ID

# clean up
# we'll keep the mp3 in tmp since that can be an expensive operation

# if wav exists
if [ -f "/tmp/$YT_ID.wav" ]; then
    rm "/tmp/$YT_ID.wav"
fi

Might require some changes to the binary path for whisper; whisper.bin is not typical; I wanted to differentiate an existing whisper script on my machine from the base, unadorned, raw whisper binary. Tweak to taste.