Whisper long file transcription + GPT4 postprocessing

transcribe.sh

#!/bin/bash

# Check if an input file was provided
if [ "$#" -ne 1 ]; then
    echo "Usage: $0 <input_audio_file>"
    exit 1
fi

# Variables
INPUT_FILE=$1
BASENAME=$(basename "$INPUT_FILE" | sed 's/\(.*\)\..*/\1/')
OUTPUT_DIR="chunks"
WHISPER_API_URL="https://api.openai.com/v1/audio/transcriptions"  # Whisper API URL
API_KEY="sk-XYZ"  # Whisper API key
MODEL="whisper-1"  # Set the model parameter
POST_PROCESS_API_URL="https://api.openai.com/v1/chat/completions"  # Post-processing API URL
SYSTEM_PROMPT="You are a helpful transcription assistant. Your task is to correct any spelling discrepancies and clean up the transcribed text. Remove filler transcription such as Napisy przygotowane przez XYZ. Only add necessary punctuation such as periods, commas, and capitalization, and use only the context provided."

# Create output directory if it doesn't exist
mkdir -p $OUTPUT_DIR

# Step 1: Split the audio file into 20MB chunks
ffmpeg -i $INPUT_FILE -f segment -segment_time 120 -c copy $OUTPUT_DIR/output%03d.wav

# Create a temporary file for the transcript
TRANSCRIPT_FILE=$(mktemp)

# Initialize an empty string for the context
context=""

# Step 2: Transcribe each chunk and merge transcriptions
for chunk in $OUTPUT_DIR/*.wav; do
    echo "Processing $chunk..." >&2
    
    # Step 2a: Prepare the prompt with the context
    prompt="\n\nTranscript so far:\n$context\n\n"

    # Step 2b: Send chunk to Whisper API with the context prompt
    response=$(curl -s -X POST -H "Authorization: Bearer $API_KEY" -H "Content-Type: multipart/form-data" \
        -F "file=@${chunk}" -F "model=$MODEL" -F "prompt=$prompt" $WHISPER_API_URL)

    # Step 2c: Extract transcription text from response
    transcription=$(echo $response | jq -r '.text')

    echo $response

    # Step 2d: Append transcription to the final transcription file
    echo "$transcription" >> $TRANSCRIPT_FILE

    # Update context with the current segment's transcription
    context="$context\n$transcription"
done

# Step 3: Post-process the full transcription text
full_transcription=$(cat $TRANSCRIPT_FILE)

# Properly escape the system prompt and transcription text for JSON
SYSTEM_PROMPT_ESCAPED=$(echo "$SYSTEM_PROMPT" | jq -sRr @json)
FULL_TRANSCRIPTION_ESCAPED=$(echo "$full_transcription" | jq -sRr @json)

# Create the JSON payload
read -r -d '' JSON_PAYLOAD <<EOF
{
  "model": "gpt-4o",
  "temperature": 0,
  "messages": [
    {"role": "system", "content": $SYSTEM_PROMPT_ESCAPED},
    {"role": "user", "content": $FULL_TRANSCRIPTION_ESCAPED}
  ]
}
EOF

# Send the post-processing request
response=$(curl -s -X POST -H "Authorization: Bearer $API_KEY" -H "Content-Type: application/json" \
    -d "$JSON_PAYLOAD" $POST_PROCESS_API_URL)
    
echo $response

# Extract the corrected text from the response
corrected_text=$(echo $response | jq -r '.choices[0].message.content')

# Output the corrected transcription to a .md file
OUTPUT_FILE="${BASENAME}.md"
echo "$corrected_text" > "$OUTPUT_FILE"
echo "Corrected transcription saved to $OUTPUT_FILE"

# Cleanup
rm -rf $OUTPUT_DIR
rm $TRANSCRIPT_FILE

echo "Transcription and post-processing complete." >&2