elevenlabs-transcribe_skill

Overview

This skill transcribes audio to text using ElevenLabs Scribe with support for batch files, realtime streams, microphone input, and local realtime playback. It handles 90+ languages, speaker diarization, event tagging, and can emit plain text or detailed JSON with timestamps. Designed for agents and pipelines that need reliable, production-ready speech-to-text.

How this skill works

The skill sends audio to the ElevenLabs Scribe API after optionally preprocessing with ffmpeg for format compatibility. It supports synchronous batch transcription, realtime streaming from URLs or the microphone, and incremental partials for live feedback. Results can be returned as plain text or full JSON that includes word-level timestamps, speaker IDs, and event tags.

When to use it

• Transcribing recorded interviews, podcasts, or meetings from local files.
• Realtime captioning for live streams, radio, or testing local playback in realtime mode.
• Voice-driven agents or tools that accept microphone input for commands or dictation.
• Generating annotated JSON (timestamps, speaker diarization) for search, indexing, or downstream NLP.
• Quiet/agent mode when you need clean stdout without status logs.

Best practices

• Install ffmpeg system-wide to ensure robust format support and avoid transcoding failures.
• Set ELEVENLABS_API_KEY in the environment before running to prevent authentication errors.
• Use --diarize for multi-speaker audio and --json when you need timestamps or structured output.
• Prefer realtime streaming with --partials for low-latency applications and user-facing feedback.
• Limit file size and duration within supported limits (up to 3GB or 10 hours) to avoid API or local resource issues.

Example use cases

• Transcribe an MP3 interview to searchable text and generate word-level timestamps for editing.
• Monitor and transcribe a live radio stream URL in realtime for clipping and highlights.
• Run microphone-based dictation in an agent with --quiet to capture user input without agent noise.
• Process meeting recordings with speaker diarization to attribute notes and action items.
• Export full JSON transcripts with event tags to detect laughter, music, or applause for content analysis.