nuva-lab/skills/chunk-process

Chunk Process Skill

Smart video chunking and MLX-accelerated transcription for long-form content.

Problem Solved

• Raw footage too long for single Gemini upload (~47 min = 5GB+)
• Need word-level timestamps for precise cutting
• Fixed-length chunks break mid-sentence

Smart Chunking

Instead of fixed 5-minute segments, smart_chunk.py finds natural break points:

python skills/chunk-process/smart_chunk.py raw_footage.mp4 -o chunks/

How it works:

1. Detect silence regions (>500ms gaps)
2. Target 2.5-3.5 minute chunks
3. Split at natural pauses, not mid-sentence
4. Output: chunk_001.mp4, chunk_002.mp4, ...

Options:

• --min-chunk 150: Minimum chunk length (seconds)
• --max-chunk 210: Maximum chunk length (seconds)
• --silence-thresh -40: Silence detection threshold (dB)

MLX Transcription

mlx_transcribe.py uses MLX-accelerated Qwen3-ASR for fast transcription on Mac:

# Single file
python skills/chunk-process/mlx_transcribe.py audio.wav -o transcript.json

# Batch process chunks
python skills/chunk-process/mlx_transcribe.py chunks/ --batch --word-timestamps

Features:

• 3-5x faster than CPU on Apple Silicon
• Auto language detection (Chinese/English/Mixed)
• Word-level timestamps (~30ms precision)
• Outputs: transcript.json with segments + words

Combined Pipeline

# 1. Smart chunk the video
python skills/chunk-process/smart_chunk.py raw.mp4 -o chunks/

# 2. Transcribe all chunks
python skills/chunk-process/mlx_transcribe.py chunks/ --batch --word-timestamps

# Output: chunks/transcript.json (merged from all chunks)

Output Format

{
  "language": "English",
  "segments": [
    {"text": "First sentence.", "start": 0.0, "end": 2.5},
    {"text": "Second sentence.", "start": 2.5, "end": 5.0}
  ],
  "words": [
    {"text": "First", "start": 0.0, "end": 0.3},
    {"text": "sentence", "start": 0.3, "end": 0.8}
  ],
  "full_text": "First sentence. Second sentence..."
}

Why Smart Chunking?

| Fixed Chunks | Smart Chunks | |--------------|--------------| | Breaks mid-word | Breaks at pauses | | 5 min arbitrary | 2.5-3.5 min natural | | Hard cuts | Clean transitions | | Timestamp gaps | Continuous timeline |