A Desktop Whisper Client That Keeps Audio Off the Internet

What it does

Vibe is a Tauri-based desktop application that runs OpenAI’s Whisper model locally on macOS, Windows, and Linux. It transcribes audio and video files, system audio, or microphone input entirely offline, then exports results to formats like SRT, VTT, PDF, or DOCX. The project also ships with a CLI and an HTTP API with Swagger documentation, so it can fit into automated workflows beyond the GUI.

The interesting bit

Instead of treating Whisper as a command-line utility, Vibe packages it into a polished desktop experience with GPU acceleration across vendors—Vulkan for AMD and Intel, CoreML on Apple Silicon—plus speaker diarization and VAD-backed stable timestamps for subtitle work. It also layers on optional summarization through the Claude API or a local Ollama instance, making it a rare hybrid of local inference and optional external AI.

Key highlights

• Fully offline transcription; no audio data leaves the device
• GPU-optimized inference via Vulkan, CoreML, and Nvidia support
• Exports to SRT, VTT, TXT, HTML, PDF, JSON, and DOCX
• Batch processing, website audio extraction, and speaker diarization
• Built-in HTTP API and CLI for headless or scripted workflows

Caveats

• Mobile support is listed as “coming soon” but not yet available
• Stable timestamp mode is explicitly marked as slower, trading speed for subtitle-grade accuracy
• Summarization relies on external services (Claude API) or a separate local Ollama setup; it is not handled by the core offline model

Verdict

Content creators and developers who need private, local transcription with flexible export options should look here; those wanting a fully managed cloud API with zero local hardware concerns will find the GPU and model management overhead unnecessary.