This tool provides an offline, local voice recognition service that converts audio/video into text, with support for JSON, SRT, and plain text output formats. It's designed for users needing to self-host a speech-to-text solution, offering accuracy comparable to OpenAI's API, and is particularly useful for developers and researchers working with audio data.

How It Works

The project leverages the faster-whisper open-source model, known for its efficiency and accuracy. It supports various model sizes (tiny to large-v3), allowing users to balance performance with computational resource requirements. The tool operates as a local web service, accessible via a browser interface or an API, and automatically utilizes NVIDIA GPU acceleration via CUDA if configured.

Quick Start & Requirements

• Installation: Download pre-compiled Windows binaries from Releases or clone the repository and install dependencies (pip install -r requirements.txt).
• Prerequisites: Python 3.9-3.11, FFmpeg (Windows users need to extract ffmpeg.exe and ffprobe.exe to the project directory). NVIDIA GPU with CUDA 11.x/12.x toolkit and cuDNN for GPU acceleration.
• Models: Download model archives and place them in the models directory.
• Running: Execute python start.py to launch the local web UI.
• Docs: English README

Highlighted Details

• Offers an API endpoint for programmatic access.
• Supports multiple languages for transcription.
• Automatic detection and utilization of CUDA for NVIDIA GPUs.
• Provides options for different output formats (JSON, SRT, text).

Maintenance & Community

• Active development with a Discord community available via invite.
• Project acknowledges dependencies on faster-whisper, Flask, and FFmpeg.

Licensing & Compatibility

• The specific license is not explicitly stated in the README, but it relies on other open-source projects with their own licenses. Compatibility for commercial use or closed-source linking should be verified.

Limitations & Caveats

• Chinese language output may sometimes be in Traditional Chinese.
• GPU acceleration requires careful setup of CUDA and cuDNN; incorrect configuration can lead to errors or crashes.
• Large models (large-v3) demand significant GPU VRAM (8GB+ recommended), and insufficient VRAM can cause crashes, especially with larger audio/video files.