Local Audio to SRT Transcription Tool

Problem

Deploying an audio-to-subtitle workflow on Apple Silicon often fails less on the model and more on project setup: cloning source, installing frontend and backend dependencies, and wiring the first local run. When network access to a repository is unstable, or when the user explicitly avoids git-based deployment, manually assembling the Audio2SRT GUI is error-prone. This skill addresses that case by embedding the project templates and generating a runnable local web tool from scratch.

How It Works

The skill targets MLX Whisper transcription and Qwen translation models. It typically confirms the destination directory, then writes configuration, backend service, frontend components, and startup scripts in dependency order. It also prepares the models/ directory and ensures the startup script is executable. On launch, the workflow downloads missing models from ModelScope, starts the Python backend and the Vite frontend, and exposes a local interface for uploading audio, transcribing it, and translating the result. The key benefit is deterministic local generation from embedded templates rather than pulling and adapting an external repository.

Boundaries

It is scoped to macOS 14.0+ on Apple Silicon and assumes a Python and Node.js environment. First run requires downloading models of roughly 4GB+, so initial startup can be slow. If the user specifically wants to clone an existing Gitee project, a deploy-style skill is more appropriate. If the task is a one-off transcription without a web interface, a simpler local CLI may be better.