Self-hosted speech-to-text that mimics OpenAI’s API

What it does

This is a local, offline transcription tool that wraps the faster-whisper inference engine in a web interface and API layer. You upload audio or video through a browser tab, and it returns plain text, JSON, or timestamped SRT subtitles. It is designed as a drop-in substitute for OpenAI’s transcription API or Baidu’s speech service, running entirely on your own hardware.

The interesting bit

Instead of exposing a generic REST endpoint, it mimics OpenAI’s API format at http://127.0.0.1:9977/v1, so existing client code can switch to local hosting with a one-line URL change. For Windows users, a precompiled binary is available that launches the server and opens the UI automatically.

Key highlights

• Runs fully offline using faster-whisper models from tiny up to large-v3
• Browser-based drag-and-drop UI plus a local REST API on port 9977
• OpenAI-compatible transcription endpoint for easy client migration
• Outputs plain text, JSON, or SRT subtitles with timestamps
• Supports 15 languages including Chinese, English, Japanese, and Arabic

Caveats

• Large models can exhaust RAM or GPU VRAM; the README warns against using large-v3 without proper NVIDIA hardware and CUDA setup, and notes that sub-8 GB cards should avoid large models on files over 20 MB
• Chinese transcription may occasionally output traditional characters instead of simplified
• Windows CUDA setup is fragile: missing cuBLAS DLLs or misconfigured cudnn can cause silent crashes or “cublasxx.dll not found” errors

Verdict

Best for developers who need a private, offline replacement for cloud transcription APIs without rewriting client code. Look elsewhere if you want managed autoscaling or a polished SaaS experience; this is a local tool that demands manual model management and careful VRAM budgeting.