asr-claw（阿瑟爪）

asr-claw

Speech recognition CLI for AI agent automation. Transcribe audio streams from stdin, files, or URLs with multiple ASR engines — local and cloud.

Triggers

• User wants to transcribe audio, speech, or voice to text
• User needs speech recognition or ASR
• User wants to convert audio/voice recordings to text
• User wants to monitor live audio / livestream speech
• User asks about 语音识别、语音转文字、转写、直播语音
• adb-claw audio capture output needs to be transcribed
• User wants subtitles (SRT/VTT) generated from audio

Binary

The asr-claw binary is located at ${CLAUDE_PLUGIN_ROOT}/bin/asr-claw.

If it does not exist, the SessionStart hook will build or download it automatically.

Setup

Quick Start (Mac)

# Install the qwen-asr engine (builds C binary + downloads 0.6B model ~1.9GB)
asr-claw engines install qwen-asr

# Verify
asr-claw engines list
asr-claw doctor

OpenClaw Setup

After installing the skill via ClawHub, configure settings:

# Set default language (default: zh)
claw config set asr-claw.default_lang en

# Use a larger model
claw config set asr-claw.model Qwen/Qwen3-ASR-1.7B

# For China users — set HuggingFace mirror
claw config set asr-claw.hf_mirror https://hf-mirror.com

# Custom model path (e.g., shared NAS)
claw config set asr-claw.model_path /mnt/models/Qwen3-ASR-0.6B

# Re-run install after changing model settings
asr-claw engines install qwen-asr

Settings are stored in ~/.asr-claw/config.yaml:

default:
  engine: qwen-asr
  lang: zh
  format: json

engines:
  qwen-asr:
    binary: ~/.asr-claw/bin/qwen-asr
    model_path: ~/.asr-claw/models/Qwen3-ASR-0.6B

Cloud Engines (no local model needed)

# OpenAI Whisper API
export OPENAI_API_KEY=sk-...
asr-claw transcribe --file audio.wav --engine openai

# Volcengine Doubao (火山引擎)
export DOUBAO_API_KEY=...
asr-claw transcribe --file audio.wav --engine doubao

# Deepgram (native streaming)
export DEEPGRAM_API_KEY=...
asr-claw transcribe --file audio.wav --engine deepgram

Commands

transcribe — Core: audio to text

# File transcription
asr-claw transcribe --file meeting.wav --lang zh

# Pipe from stdin
cat audio.wav | asr-claw transcribe --lang zh

# Streaming (real-time, from adb-claw or ffmpeg)
adb-claw audio capture --stream --duration 60000 | asr-claw transcribe --stream --lang zh

# Subtitle output
asr-claw transcribe --file lecture.wav --format srt > lecture.srt
asr-claw transcribe --file lecture.wav --format vtt > lecture.vtt

# Specify engine
asr-claw transcribe --file audio.wav --engine whisper --lang en

Flags:

engines — Manage ASR engines

asr-claw engines list                    # List all engines + status
asr-claw engines install qwen-asr       # Install local engine (Mac)
asr-claw engines info qwen-asr          # Engine details
asr-claw engines start qwen3-asr        # Start vLLM service engine
asr-claw engines stop qwen3-asr         # Stop service engine
asr-claw engines status                  # Running engines

doctor — Environment check

asr-claw doctor    # Check platform, engines, dependencies

Engine Matrix

Output Format

All commands output JSON envelope:

{
  "ok": true,
  "command": "transcribe",
  "data": {
    "segments": [{"index": 0, "start": 0.0, "end": 2.5, "text": "..."}],
    "full_text": "...",
    "engine": "qwen-asr",
    "audio_duration_sec": 5.5
  },
  "duration_ms": 1230,
  "timestamp": "2026-03-13T10:00:00Z"
}

Use -o text for plain text, -o quiet for silent.

With adb-claw

# Real-time transcription from Android device
adb-claw audio capture --stream --duration 60000 | asr-claw transcribe --stream --lang zh

# Record then transcribe
adb-claw audio capture --duration 30000 --file recording.wav
asr-claw transcribe --file recording.wav --lang zh

# Save audio + transcribe simultaneously
adb-claw audio capture --stream --duration 0 | tee backup.wav | asr-claw transcribe --stream