asr-skills（紧急救援技能）

ASR Transcription Skill

Provide local audio/video transcription with speaker diarization, multiple output formats, and progress indication.

Purpose

Enable users to transcribe audio and video files to text with automatic speaker identification, supporting multiple subtitle formats while preserving privacy through local processing.

When to Use

This skill triggers when the user:

• Wants to transcribe an audio file (MP3, WAV, M4A, FLAC)
• Wants to transcribe a video file (MP4, AVI, MKV)
• Needs subtitles or captions generated from media
• Wants to identify different speakers in audio
• Needs timestamped transcription output

Quick Start

Basic Transcription

# Transcribe audio file (outputs TXT by default)
python3 skills/asr/scripts/transcribe.py path/to/audio.mp3

# Transcribe video file
python3 skills/asr/scripts/transcribe.py path/to/video.mp4

Output Formats

python3 skills/asr/scripts/transcribe.py audio.mp3 -f json   # Structured JSON with metadata
python3 skills/asr/scripts/transcribe.py audio.mp3 -f srt    # SubRip subtitles
python3 skills/asr/scripts/transcribe.py audio.mp3 -f ass    # ASS/SSA subtitles with speaker styling
python3 skills/asr/scripts/transcribe.py audio.mp3 -f md     # Markdown with speaker sections

Python API

from asr_skill import transcribe

result = transcribe("meeting.mp4", format="srt")
print(f"Output: {result['output_path']}")
print(f"Speakers: {result.get('speakers', [])}")

Asynchronous Execution (Recommended for Long Files)

Avoid timeouts by running transcription in the background:

# Start async task
python3 skills/asr/scripts/transcribe.py long_video.mp4 --async
# Output: {"task_id": "a1b2c3d4", "status": "queued", ...}

# Check status
python3 skills/asr/scripts/transcribe.py --status a1b2c3d4
# Output: {"task_id": "a1b2c3d4", "status": "processing", "progress": 45, ...}

# List recent tasks
python3 skills/asr/scripts/transcribe.py --list

Core Features

Speaker Diarization

Automatically identifies and labels different speakers:

• Speaker A, Speaker B, Speaker C, etc.
• Per-segment timestamps
• Overlap detection marked with [OVERLAP]

Hardware Auto-Detection

Detects and uses the best available hardware:

• CUDA GPU (NVIDIA)
• Apple MPS (Apple Silicon)
• CPU fallback with notification

Long Audio Support

Handles audio files longer than 1 hour:

• VAD-based intelligent segmentation
• Memory-efficient processing
• Progress indication during transcription

Multiple Output Formats

Implementation Details

Processing Pipeline

1. Input validation - Check file exists and format supported
2. Hardware detection - Auto-detect GPU/MPS/CPU
3. Video extraction - Extract audio from video files via FFmpeg
4. Audio preprocessing - Resample to 16kHz mono
5. Model loading - Load FunASR models (cached locally)
6. Transcription - Run ASR with speaker diarization
7. Formatting - Output in requested format
8. Cleanup - Remove temporary files

Model Components

• ASR Model: Paraformer-large (Chinese optimized)
• VAD Model: FSMN-VAD (voice activity detection)
• Punctuation: CT-Transformer
• Speaker: CAM++ (speaker diarization)

File Locations

• Models cached in: ./models/
• Output defaults to: same directory as input
• Temp files: auto-cleaned after processing

Troubleshooting

Common Issues

"FFmpeg not found"

• FFmpeg auto-installed via imageio-ffmpeg
• Check internet connection for first run

"CUDA out of memory"

• System falls back to CPU automatically
• Try shorter audio segments

"No speakers detected"

• Speaker diarization requires multi-speaker audio
• Single speaker audio shows "Speaker A" only

Additional Resources

Reference Files

For detailed format specifications:

• references/output-formats.md - Complete format documentation

Scripts

Utility scripts for batch processing:

• scripts/transcribe.py - Batch transcription script

Examples

Working examples:

• examples/basic_usage.py - Python API examples
• examples/cli_examples.sh - CLI usage examples

Requirements

• Python >= 3.10
• FunASR (auto-installed)
• FFmpeg (auto-installed via imageio-ffmpeg for video)

Notes

• First run downloads models (~1GB total)
• All processing happens locally for privacy
• Chinese language optimized for v1