asr-skill

This skill should be used when the user asks to "transcribe audio", "transcribe video", "convert speech to text", "generate subtitles", "create captions", "i...

Security Scan

The skill is internally coherent for a local ASR/transcription use case and does not request credentials or external installs in its manifest, but there are implementation details you should review (background worker, local task files, and the model-download behavior) before installing in a sensitive environment.

Purpose & Capability

Name, description, SKILL.md, examples, and scripts consistently describe a local ASR transcription tool with speaker diarization and multiple output formats. The code present (transcribe.py and examples) implements task management, async/background execution, duration estimation, and calls into an asr_skill Python API for the heavy work, which fits the stated purpose.

Instruction Scope

Instructions and code operate on local files (audio/video), extract audio with ffprobe/ffmpeg, resample, load models, and write outputs. The script creates a .asr_skill directory and tasks.json in the project root to track async jobs and spawns a detached background worker (subprocess.Popen with start_new_session=True and stdout/stderr redirected to DEVNULL). There is no evidence in the provided files of external network exfiltration, but the SKILL.md notes first-run model downloads (~1GB) and those download routines are not present in the reviewed files (they likely live in the missing asr_skill package), so you should verify where models are fetched from if local-only processing and privacy are essential.

Install Mechanism

There is no install spec (instruction-only), which minimizes installer risk. However, SKILL.md claims auto-install of FunASR and FFmpeg via imageio-ffmpeg and that models are downloaded on first run; those actions are not in the shown script and would occur in imported modules (not included here). Because downloads/extracts are plausible but not visible, verify the code that performs automatic installs/downloads before running.

Credentials

The skill declares no required environment variables, credentials, or config paths and the provided code does not attempt to read secrets or unrelated system configuration. It uses local filesystem storage for tasks and relies on available binaries (ffprobe) or Python libraries (librosa), which is proportional to the stated functionality.

Persistence & Privilege

The skill does not set always:true and does not modify other skills. It does persist local state by creating a .asr_skill directory and tasks.json under the project root and spawns detached background worker processes for async jobs. These behaviors are expected for async transcription but can create persistent files and background processes that you may want to control or audit.

Assessment

This skill appears to implement what it claims (local transcription with diarization) and doesn't request secrets. Before installing or running it: 1) Inspect the asr_skill package (not included in the reviewed files) to see where and from which hosts models are downloaded and whether any network calls are made; 2) Be aware the script will create a hidden .asr_skill/tasks.json in the project root and will spawn detached background worker processes (their stdout/stderr are suppressed), so run in an isolated environment if you need to restrict resource or process creation; 3) Ensure dependencies (FunASR, imageio-ffmpeg/FFmpeg, librosa, etc.) are installed from trusted sources or pinned to known versions; 4) If privacy is critical, confirm model downloads occur from trusted hosts and consider auditing network activity on first run. If you want higher assurance, ask the publisher for the full asr_skill package source or run the skill in a sandboxed environment first.

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

Format	Extension	Use Case
txt	.txt	Plain text with timestamps
json	.json	Structured data with word-level info
srt	.srt	Video subtitles
ass	.ass	Styled subtitles
md	.md	Documentation with speaker sections

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