a0_transcribbler plugin thumbnail

a0_transcribbler

Voice messages become text conversations! Automatically transcribe audio attachments, URLs, and YouTube videos before LLM processing.

Author neurocis 1 stars Version 2.0.0 Updated
VoiceMultimediaTranscriptionSpeech To Text
GitHub Repository Discussion How to install

README

a0_transcribbler

📣 Voice messages become text conversations! This plugin enables users to send voice memos via Telegram, Signal, WebUI, or any input channel and have them automatically transcribed and processed by the LLM as if they were typed text.

No more typing long messages — just speak and the AI understands.

Automatic audio transcription plugin for Agent Zero. Transcribes audio file attachments, audio URLs, and YouTube video audio before LLM processing, injecting the transcription text directly into the conversation context.

Features

  • Voice Memo Transcription — Send voice messages via Telegram, Signal, or any platform and they're automatically converted to text for the LLM. Perfect for hands-free interaction and accessibility.

  • Audio File Transcription — Automatically detects and transcribes audio attachments (voice messages, audio files) sent via Telegram, WebUI, or any other input channel. Supports .ogg, .mp3, .wav, .m4a, .opus, .flac, .aac, .wma, .webm, and more.

  • Audio URL Transcription — Scans URLs in messages and probes their MIME type via HTTP HEAD request. URLs serving audio/* content are automatically downloaded and transcribed. YouTube URLs are handled separately with dedicated subtitle-first logic.

  • YouTube Video Transcription — Detects YouTube URLs in messages, first attempts to fetch existing subtitles/captions, then falls back to downloading the audio track via yt-dlp and transcribing with Whisper. Includes configurable duration limits to prevent excessively long downloads.

  • Pre-LLM Injection — Transcriptions are prepended to the user message before the LLM sees it, so the agent can reason about audio content naturally.

  • Reuses Existing STT — Built on Agent Zero's built-in OpenAI Whisper engine. No additional STT models or API keys required.

How It Works

  1. When a user sends a message, the plugin intercepts it via the hist_add_user_message extensible hook (before history is written).
  2. It scans attachments for audio file extensions.
  3. It scans the message text for non-YouTube URLs and probes each with an HTTP HEAD request to detect audio/* MIME types.
  4. It scans the message text for YouTube URLs.
  5. Detected audio is converted to WAV (16kHz mono) via ffmpeg and transcribed using the Whisper model configured in Agent Zero settings.
  6. Transcription text is prepended to the original message with a clear label.
  7. The LLM receives the enriched message and can respond to the audio content.

Requirements

  • ffmpeg — Pre-installed in Agent Zero Docker images.
  • OpenAI Whisper — Pre-installed in Agent Zero (openai-whisper package).
  • yt-dlp — Installed automatically by the plugin's hooks.py on first install. Required only for YouTube transcription.

Configuration

All settings are accessible from the Agent Zero Plugin Settings UI:

Setting Default Description
audio_transcription_enabled true Enable/disable audio file transcription
youtube_transcription_enabled true Enable/disable YouTube URL transcription
url_audio_transcription_enabled true Enable/disable audio URL MIME-type scanning
youtube_max_duration 3600 Max YouTube video length in seconds (0 = no limit)
url_audio_max_size 50 Max audio URL download size in MB
audio_extensions see below List of audio file extensions to detect
transcription_label [Transcription] Label prefix for transcription blocks
include_filename true Include filename in transcription header
model_bypass_enabled true Skip audio transcription for audio-capable models
audio_capable_models gemini-*, gpt-4o-audio-*, ... Model name wildcard patterns

Installation

  1. Place the a0_transcribbler folder in usr/plugins/.
  2. Enable the plugin in the Agent Zero Plugins UI.
  3. The plugin will auto-install yt-dlp on first activation.

Example Output

When a user sends a voice message saying "Remind me to buy groceries", the LLM receives:

[Transcription] (voice_abc123.ogg):
Remind me to buy groceries

---

[Telegram message from user]
[Voice message — see attachment]

Version

2.0.0