菜单
ASR with ~30ms timestamp precision using Qwen3-ASR + ForcedAligner
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
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)
Text-guided audio source separation using SAM-Audio via mlx-audio
| name | transcribe-audio |
| description | ASR with ~30ms timestamp precision using Qwen3-ASR + ForcedAligner |
Transcribe audio/video files with precise timestamps using Qwen3-ASR and Qwen3-ForcedAligner. Runs on CPU for maximum quality.
| Model | Purpose | Precision |
|---|---|---|
| Qwen3-ASR-1.7B | Speech recognition (52 languages) | SOTA accuracy |
| Qwen3-ForcedAligner-0.6B | Timestamp alignment | ~30ms |
# Full transcription with timestamps (default)
python skills/transcribe-audio/transcribe.py audio.wav
# Save to file
python skills/transcribe-audio/transcribe.py audio.wav --output captions.json
# Specify language (auto-detects if not specified)
python skills/transcribe-audio/transcribe.py audio.wav --language Chinese
# Fast mode (no timestamps)
python skills/transcribe-audio/transcribe.py audio.wav --no-timestamps
# Align existing text to audio (ForcedAligner only)
python skills/transcribe-audio/transcribe.py audio.wav --align-text "Your transcript text here"
Compatible with Remotion captions.json:
{
"segments": [
{"text": "当", "startMs": 0, "endMs": 150},
{"text": "全", "startMs": 150, "endMs": 280},
{"text": "世界", "startMs": 280, "endMs": 520},
...
],
"full_text": "当全世界都在追AI的时候...",
"language": "Chinese",
"model": "Qwen3-ASR-1.7B",
"aligner": "Qwen3-ForcedAligner-0.6B"
}
Transcribes audio and aligns timestamps:
python transcribe.py audio.wav
Use when you already have the transcript:
python transcribe.py audio.wav --align-text "已知的文字内容"
This skips ASR and uses ForcedAligner directly for ~30ms precision.
from transcribe import transcribe_audio, transcribe_and_align
# Full transcription with timestamps
result = transcribe_audio("audio.wav", language="Chinese")
print(result["segments"]) # Word-level timestamps
# Align existing text
result = transcribe_and_align("audio.wav", text="已知的文字", language="Chinese")
print(result["segments"]) # ~30ms precision timestamps