Whisper is OpenAI’s open-source speech-recognition model. With a local Whisper provider, AstrBot can turn a voice message into text before sending it to your chat model. Local processing avoids a per-minute transcription API bill, though it still uses your own hardware.
Install Whisper in AstrBot’s environment
The original dashboard offers a Python-package installation option under platform logs. Install openai-whisper there, or use the Python environment that runs AstrBot:
pip install -U openai-whisperWhisper also requires FFmpeg. On Debian or Ubuntu:
sudo apt update
sudo apt install ffmpegThe initial model download can take time. Check the Whisper repository if package installation fails.
Add the local provider
Open Model Providers, choose speech-to-text, and add Whisper (Local). Give it an ID such as whisper_selfhost and choose a model name.
The main choices are tiny, base, small, medium, large, and turbo. Tiny is a practical starting point on a CPU-only device; larger models need more memory and usually more compute. Turbo is an optimized large-v3 variant designed for faster transcription with a modest accuracy tradeoff.
Speed and accuracy depend on hardware, language, and audio quality. The published speed comparisons are benchmark results, not a guarantee that every large-model transcription is exactly ten times slower. Turbo is not intended for translating non-English speech into English.
Enable voice input
Save the provider, then enable speech-to-text in the bot’s general configuration and select it. Send a short voice message and compare the transcript with what you said. If you prefer a hosted transcription API, check its current model pricing; the original post quoted $0.006 per minute for whisper-1.
Adapted from the original Chinese article, published on March 15, 2026.
Comments NOTHING