Super-Transcribe

A speech-to-text skill with two bundled engines for the best speed and featureset — NVIDIA Parakeet (NeMo) for the former, and faster-whisper (CTranslate2) for the latter.

The setup script installs the fastest backend for your setup, and the main script auto-selects the best backend for your task (and loads the other backend if necessary).

First-Time Setup

If you set this up with your agent, it will use the setup script, which detects your platform, GPU, Python version, and optional dependencies. Next, it will install and configure the optimal backend(s).

See Prerequisites below for more information about backend compatibility.

Prerequisites

Why Two Backends?

In general, Parakeet has better speed and accuracy, while faster-whisper has more features.

The router automatically picks the most suitable backend for your task. The agent can override this with --backend parakeet or --backend faster-whisper.

Features (Selected)

Shared (both backends):

• 10 output formats
• Speaker diarization
• Chapter detection
• Filler removal
• Denoise/normalize audio
• Check RSS/podcast feeds
• Batch processing

Parakeet-only (selected features):

• --fast (110M model)
• --multitalker (overlapping speech)
• Canary translation (allowing translation between European languages, not just translating to English like in faster-whisper)

faster-whisper-only (selected features):

• --translate (to English only)
• --initial-prompt (makes decoder focus on transcribing certain words correctly)
• --hotwords (like --initial-prompt but more focused on getting the exact spellings right)
• --multilingual (supports multiple languages in the same file)
• 99+ languages

Agent Integration

This skill is primarily designed for OpenClaw agents.

Key features to aid agent use:

• --probe — Check audio duration/format before committing to transcription

  # Probe before deciding to transcribe
  ./scripts/transcribe --probe recording.mp3
  # → {"file":"recording.mp3","duration":2714.5,"duration_human":"45m 14s",...}
  

• --agent — Compact JSON output with text, duration, language, confidence, speaker info, and summary hints

  # Agent mode with file output
  ./scripts/transcribe --agent -o /tmp/transcript.txt audio.ogg
  # → {"text":"...","duration":4.2,"avg_confidence":0.94,"summary_hint":{"first":"...","last":"..."},...}
  

• Exit codes — 0 (success), 1 (error), 2 (missing dep), 3 (bad input), 4 (GPU OOM)

• First-run messaging — Notifies the agent when a backend is setting up for the first time

Output Formats

text (default) · json · srt · vtt · ass · lrc · ttml · csv · tsv · html

Requirements

• Python 3.10+
• NVIDIA GPU + CUDA (highly recommended — CPU is 50-100× slower)
• Optional: ffmpeg (for non-WAV input, preprocessing, burn-in)
• Optional: yt-dlp (for YouTube/URL input)

Platform Support

Documentation

• SKILL.md — Full reference with all options, model tables, and agent guidance

Licenses

• Parakeet TDT v3: CC-BY-4.0
• faster-whisper: MIT