whisperkit-cli

On-device Whisper transcription + Qwen3-TTS synthesis
Local file-based audio I/O -- models are downloaded from HuggingFace on first run, then all inference runs on-device with no network required. Perfect for agents that receive voice messages/attachments and reply with text or generated audio files.

The agent saves incoming audio attachments to a temp path, runs the CLI, and either returns the transcribed text in chat or attaches the generated .wav/.m4a reply.

Why agents love this skill

• Runs on ANE -- no GPU contention, low power, always available
• No API keys, no per-request costs, no data leaves the machine after setup
• One-time model download on first run, then fully offline
• Handles audio files from user messages (m4a, wav, mp3, flac)
• Generates reply audio files the agent can attach/send
• 9 built-in voices + 10 languages
• Natural-language style instructions (1.7B model)

Installation