菜单
Text-guided audio source separation using SAM-Audio via mlx-audio
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | separate-audio |
| description | Text-guided audio source separation using SAM-Audio via mlx-audio |
Isolate specific sounds from audio using natural language text prompts. Uses Meta's SAM-Audio model via mlx-audio for native Mac M2/M3 inference.
# Extract speaker by description
python skills/separate-audio/separate.py panel.wav --prompt "man speaking" --output speaker.wav
# Extract with time hint
python skills/separate-audio/separate.py video.mp4 --prompt "applause" --span 10.5-12.0
# Save both target and residual
python skills/separate-audio/separate.py audio.wav --prompt "woman singing" --save-residual
| Use Case | Prompt Example |
|---|---|
| Extract single speaker | "man speaking about investments" |
| Remove background music | Separate, keep residual |
| Isolate applause | "audience applause" |
| Clean panel discussion | Run multiple times with different prompts |
from separate import separate_audio
result = separate_audio(
audio_path="panel.wav",
prompt="man speaking about space",
output_path="speaker.wav",
span=(10.5, 12.0), # Optional time hint
)
print(result["target_path"])
pip install mlx-audioThis skill is implemented but not extensively tested in the main video pipeline. The primary audio workflow uses Qwen3-ForcedAligner for caption alignment. SAM-Audio is available for advanced use cases like:
Generate karaoke-style word-level timestamps by aligning script text to audio using Qwen3-ForcedAligner + jieba for Chinese word segmentation. Use when the user says 'align captions', 'karaoke timestamps', 'word timestamps', 'caption alignment', 'sync text to audio'.
Analyze raw video content using Gemini to identify speakers, topics, key moments, and potential clip opportunities
Audio processing utilities - noise reduction, normalization, enhancement
Extract a video segment using FFmpeg with precise start/end times
Find naturally clean, coherent video segments worth keeping (selection over repair)
ASR with ~30ms timestamp precision using Qwen3-ASR + ForcedAligner