Fun-ASR: Audio Transcription

Transcribe audio files using Alibaba Cloud DashScope's Fun-ASR non-real-time speech recognition model. Supports speaker diarization, multi-language recognition, and multiple output formats (plain text, JSON, SRT subtitles).

Workflow

Audio → S3 → Fun-ASR async → Poll → Save file → Agent reads

Requirements

Environment Variables

Set these in ~/.wmyskills/.env (shared across skills), or scripts/.env (takes priority). Use scripts/.env.example as the template — add the variables to ~/.wmyskills/.env, never overwriting an existing file:

The script loads .env from ~/.wmyskills/.env and scripts/.env automatically (script directory takes priority).

Python Dependencies

pip install boto3 requests python-dotenv

Optional: ffmpeg

ffmpeg/ffprobe is needed for:

• Detecting audio channel count
• Converting multi-channel audio to mono (required for speaker diarization)

Without ffmpeg, the script skips mono conversion and uses the original file.

Usage

python fun_asr_cli.py <audio-file> [options]

Run from the scripts/ directory or provide the full path to funasrcli.py.

Positional Arguments

Options

Examples

# Basic transcription (text format, Chinese)
python fun_asr_cli.py meeting.mp3

# JSON output with full metadata (timestamps, speaker IDs, confidence)
python fun_asr_cli.py interview.wav --format json

# SRT subtitle output
python fun_asr_cli.py lecture.mp3 --format srt

# Disable speaker diarization
python fun_asr_cli.py meeting.mp3 --no-diarization

# Specify language and model
python fun_asr_cli.py japanese_audio.mp3 --language ja --model paraformer-v2

# Custom output path
python fun_asr_cli.py audio.wav --output ~/Desktop/transcript.txt

Output Formats

text (default)

Plain text with speaker labels and timestamps: [Speaker N] HH:MM:SS - HH:MM:SS, followed by the text.

json

Full JSON with all metadata — timestamps, speaker IDs, confidence scores, and the raw API response.

srt

Standard SRT subtitle format: NN \n HH:MM:SS,mmm --> HH:MM:SS,mmm \n [SN] text.

Execution Rule

Run the user's requested command directly without pre-checking dependencies, environment variables, or configuration. If something is wrong, the script will fail with a clear error — check and fix only then.

Agent Instructions

After Transcription Completes

Transcription output is saved to a file — the text content is not printed to stdout. The only line printed to stdout is Output file: <path>, which tells you where the result file was saved. The default output directory is ~/.wmyskills/fun-asr/outputs/.

When transcription succeeds, stdout contains only: Output file: <path>. Info and error logs go to stderr.

On success you must deliver the result to the user:

1. Parse the Output file: path from stdout
2. Read and present the file content to the user
3. Provide a brief summary: audio duration, number of speakers detected (if diarization was enabled), and a concise overview of the content

Error Scenarios

The script exits with specific codes for programmatic handling:

• Code 2 — Configuration error: missing env vars. Also used by argparse for invalid arguments (e.g. unknown flags, invalid model name).
• Code 3 — File error: file not found, too large, or exceeds limits.
• Code 4 — API/network error: submission or polling failed.
• Code 6 — Task failure: ASR API returned an error. Check the error message.
• Code 7 — Timeout: transcription took longer than 30 minutes.

Important Rules

• No valid speech detected: If the API returns ASRRESPONSEHAVENOWORDS or SUCCESSWITHNOVALIDFRAGMENT, the audio contains no detectable human speech (silence, too quiet, or pure noise). Stop immediately and report this to the user. Do NOT retry automatically, switch models, or modify audio parameters. Only retry if the user explicitly asks.
• Speaker diarization requires mono audio ≤ 2 hours. The script auto-converts multi-channel audio via ffmpeg. If the file exceeds 2 hours, diarization is disabled automatically.
• Diarization cost: The script charges per audio second regardless of diarization; there is no extra fee for enabling it.
• S3 cleanup: Temporary files on S3 are deleted automatically after transcription unless --keep-s3 is passed.
• Pricing: Fun-ASR costs CNY 0.00022 / second of audio. The script estimates cost before starting and reports actual cost on completion.