CastPolish

A local, open-source alternative to Auphonic for podcast and audio post-production. CastPolish runs entirely on your Mac — no cloud, no subscription, no audio ever leaves your machine.

---

What it does

• Audio processing — highpass filter, compression, limiting, and two-pass EBU R128 loudness normalization (targets −16 LUFS, podcast standard)
• Noise reduction — optional ffmpeg afftdn FFT noise reduction (checkbox in UI)
• Transcription — faster-whisper with word-level timestamps, exported as HTML, WebVTT captions, and Auphonic-compatible JSON
• AI shownotes — chapter titles, long summary, brief summary, and suggested tags via Ollama (local LLM, no API key needed)
• Speaker diarization — optional, via pyannote.audio (requires HuggingFace token)
• macOS app — one-click .app launcher with Dock icon

---

Requirements

Dependency	Install
Python 3.10+	python.org
ffmpeg	brew install ffmpeg
Ollama (optional, for shownotes)	ollama.com

---

Quick start

# 1. Clone the repo
git clone https://github.com/abc3-Mac/castpolish.git
cd castpolish

# 2. Install Python dependencies
python3 castpolish.py setup

# 3. Start the server
python3 castpolish.py serve

# 4. Open http://localhost:8765 in your browser

---

macOS App (optional)

Build a native .app with a Dock icon that launches the server with a double-click:

python3 create_macos_app.py

Then drag CastPolish.app to your Applications folder or Dock.

The app bundles castpolish.py internally and works from any location. It does not require the source folder to remain in place after the app is built.

---

Command-line usage

Start the web server

python3 castpolish.py serve
python3 castpolish.py serve --port 9000
python3 castpolish.py serve --output-dir ~/Desktop/podcast-output

Process a file directly (no browser needed)

python3 castpolish.py process "episode.mp3"

python3 castpolish.py process "episode.mp3" \
  --model small \          # tiny | base | small | medium | large-v2
  --format mp3 \           # mp3 or mp4 (aac)
  --language en \
  --task transcribe \      # transcribe | translate (translate → English)
  --lufs -16 \             # target loudness in LUFS
  --title "Episode Title" \
  --output-dir ~/my-output \
  --denoise \              # enable ffmpeg afftdn noise reduction
  --no-normalize           # skip loudness normalization

Batch processing

for f in ~/Podcasts/*.mp3; do
    python3 /path/to/castpolish.py process "$f" --model small
done

---

Output files

Each processed file gets its own folder named after the audio file:

~/CastPolish-output/
  my-episode/
    my-episode.mp3          # processed audio
    my-episode.html         # transcript with chapters, summaries, audio player
    my-episode.vtt          # WebVTT captions (YouTube-compatible)
    my-episode.json         # Auphonic-compatible transcript JSON

If you process the same file twice, folders are named my-episode-01, my-episode-02, etc.

---

AI shownotes (Ollama)

Install Ollama and pull a model:

brew install ollama
ollama pull llama3.2

Then enable shownotes in the web UI Settings tab, or pass --title "Episode Title" on the CLI. CastPolish will generate:

• Chapter titles with timestamps
• Long summary (~700 tokens)
• Brief summary (~300 tokens)
• Suggested tags

Ollama runs locally — no data leaves your machine.

---

Speaker diarization (optional)

Requires a free HuggingFace account and accepting the pyannote model license:

pip install pyannote.audio torch

Enter your HuggingFace token in Settings and enable diarization in the UI.

---

Performance guide

Hardware	tiny	small	medium
Apple Silicon (M1/M2/M3)	~0.1× realtime	~0.3× realtime	~0.8× realtime
Intel Mac (last gen)	~0.5× realtime	~1.5× realtime	~4× realtime

"0.3× realtime" means a 60-minute file takes ~18 minutes. Apple Silicon uses Core ML acceleration.

For batch jobs on Intel, --model small is the best balance of speed and accuracy.

---

Configuration

Settings are saved to ~/.castpolish/config.json. You can also configure via the Settings tab in the web UI:

• Output directory
• Default Whisper model
• Default audio format
• Target loudness (LUFS)
• Ollama host and model
• HuggingFace token

---

License

MIT — see LICENSE.

---

Acknowledgements

• faster-whisper — OpenAI Whisper via CTranslate2
• ffmpeg — audio processing and noise reduction
• Ollama — local LLM inference
• Flask — local web server
• Inspired by Auphonic