Real-time Transcription Skill

Capture any audio, get a structured summary. Real-time transcription powered by SenseVoice/FunASR.

Features

• Real-time transcription — stream audio from system (BlackHole) or microphone
• Auto summary — on stop, generate title + structured summary
• Date-based archival — results saved to archive/YYYY/MM/DD-HHMM-title.md
• Idle detection — auto-stops after 60s of silence (configurable)

Skill Location

All files are in ~/.openclaw/skills/realtime-transcription/:

realtime-transcription/
├── SKILL.md                 # This file
├── realtime_asr.py          # Background transcription process
├── summary_prompt.py        # LLM prompt builder & response parser
├── archiver.py              # Markdown archival module
├── references/
│   └── module-reference.md  # Module API reference
├── .tmp/                    # Runtime temp files
└── archive/                 # Archived outputs

Prerequisites

Python Dependencies

pip3 install sounddevice librosa funasr torch numpy

Or use the built-in installer with progress output:

cd ~/.openclaw/skills/realtime-transcription
python3 realtime_asr.py --install-deps

System Audio (optional, macOS)

For macOS system audio capture, install BlackHole: brew install blackhole-2ch

ASR Model

Download the SenseVoice model: modelscope download --model gongjy/SenseVoiceSmall --local_dir ./model/SenseVoiceSmall

Quick Start

Check Dependencies

cd ~/.openclaw/skills/realtime-transcription
python3 realtime_asr.py --check-deps

Expected output:

✅ 所有依赖已安装。
   sounddevice — PyAudio binding for microphone/system audio capture
   librosa — Audio resampling and preprocessing
   funasr — SenseVoice ASR model framework
   torch — PyTorch deep learning runtime
   numpy — Numerical array processing

If dependencies are missing, run python3 realtime_asr.py --install-deps to install them one by one with progress output.

Start Transcription

System audio (BlackHole):

cd ~/.openclaw/skills/realtime-transcription
python3 realtime_asr.py --source blackhole

Microphone:

cd ~/.openclaw/skills/realtime-transcription
python3 realtime_asr.py --source mic

With custom idle timeout (5 minutes):

cd ~/.openclaw/skills/realtime-transcription
python3 realtime_asr.py --source mic --idle-timeout 300

Disable idle timeout:

cd ~/.openclaw/skills/realtime-transcription
python3 realtime_asr.py --source mic --idle-timeout 0

Stop Transcription

Press Ctrl+C in the terminal, or:

kill $(cat .tmp/asr.pid 2>/dev/null) 2>/dev/null; rm -f .tmp/asr.pid

After Stopping — Summary & Archive

1. Read the transcript: cat .tmp/transcript.txt
2. Build the LLM prompt:

```bash

cd ~/.openclaw/skills/realtime-transcription

python3 -c "

from summary_prompt import build_summary_prompt

print(build_summary_prompt(open('.tmp/transcript.txt').read()))

"

```

1. Send the prompt to yourself (the LLM) to generate TITLE + SUMMARY
2. Parse and archive:

```bash

cd ~/.openclaw/skills/realtime-transcription

python3 -c "

from summary_prompt import parse_summary_response

from archiver import archive

transcript = open('.tmp/transcript.txt').read()

result = parse_summary_response('YOUR_LLM_RESPONSE_HERE')

path = archive(transcript, result['title'], result['summary'], 'blackhole')

print(f'Archived to: {path}')

"

```

CLI Reference

Trigger Words

Output Format

Transcript (.tmp/transcript.txt)

[14:30:00] 你好今天我们来讨论一下AI的发展
[14:30:05] AI技术在各个领域都有广泛应用

Archive (archive/YYYY/MM/DD-HHMM-title.md)

---
title: "AI发展趋势讨论"
date: 2025-05-16
time: "14:30 - 14:38"
source: blackhole
duration: 8m
---

## 摘要

- AI在医疗、金融、教育领域广泛应用
- 未来将更智能和普及

## 完整转录

[14:30:00] 你好今天我们来讨论一下AI的发展
...

Error Handling

Exit Codes