Transcribe Via Faster Whisper

Transcribe via Faster Whisper

Overview

Free, local transcription using faster-whisper running in Docker. Extracts audio from video, sends it to the local whisper server, and produces both full text and timestamped JSON output.

Prerequisites

• Docker (for faster-whisper transcription service)
• ffmpeg (for audio extraction)

Workflow

Step 1: Set Up Transcription Service

Check if a compose.yaml exists in the project with a faster-whisper service. If not, create one:

services:
  faster-whisper:
    image: fedirz/faster-whisper-server:latest-cpu
    container_name: faster-whisper
    ports: ["8000:8000"]
    volumes:
      - ./hf_cache:/root/.cache/huggingface
    restart: unless-stopped

Add hf_cache/ to .gitignore. Start with docker compose up -d.

Step 2: Extract Audio

Extract audio from video to reduce upload size (skip if input is already audio):

ffmpeg -i "$INPUT" -vn -acodec libmp3lame -q:a 4 /tmp/audio.mp3 -y

Step 3: Transcribe

1. Wait for whisper server health check:

   curl -s http://localhost:8000/health
   

2. Send for transcription with timestamps:

   curl -s http://localhost:8000/v1/audio/transcriptions \
     -F "file=@/tmp/audio.mp3" \
     -F "model=Systran/faster-whisper-small" \
     -F "response_format=verbose_json" \
     -o transcription.json
   

3. Save both the full text and the timestamped JSON.

Step 4: Output

Produce two files in the output directory:

• transcription.json — timestamped segments with start/end times
• transcription.txt — full plain text transcript

Tips

• First run downloads the model (~500MB) — subsequent runs are fast.
• For long recordings (>1h), transcription may take several minutes.
• The verbose_json format includes per-segment timestamps, which are essential for downstream tools like video cutting.
• If the server is not responding, check docker compose logs faster-whisper.