Speech Recognition Channel: STT Local Recognition Tool

What is this?

STT is an offline local speech recognition tool based on the Faster-Whisper open-source model, capable of converting human speech in video or audio to text.

Key features:

• Fully offline — no internet required
• Multiple output formats: JSON, SRT subtitles, plain text
• Built-in base model with additional models available for download
• CUDA GPU acceleration support
• Accuracy comparable to OpenAI's official API

GitHub repository: https://github.com/jianchang512/stt

Available Models

STT supports the following Whisper models, listed from lowest to highest recognition quality:

Model	Parameters	VRAM	Quality	Use Case
base	74M	~1 GB	Basic	Quick testing, simple scenarios
small	244M	~2 GB	Good	Daily use
medium	769M	~5 GB	Very good	Multilingual recognition
large	1550M	~10 GB	Excellent	High-accuracy recognition
large-v3	1550M	~10 GB	Best	Highest accuracy requirements

Model download: https://github.com/jianchang512/stt/releases/tag/0.0

Model selection tip: If GPU VRAM is limited, start with the base or small model. Larger models are more accurate but require more hardware resources.

Pre-compiled Windows Version

Requirements

• Windows 10/11 64-bit
• For GPU acceleration: NVIDIA GPU with CUDA properly configured

Installation Steps

1. Download the pre-compiled package Visit the Releases page and download the latest pre-compiled package.

2. Extract the files Extract to any location. Paths without Chinese characters or spaces are recommended, such as E:/stt.

3. Start the service Double-click start.exe and wait for the browser to automatically open the local web interface.

4. Recognize speech

   • Click the upload area or drag an audio/video file onto it
   • Select the spoken language (Chinese/English/Auto-detect, etc.)
   • Select the output format (JSON/SRT/plain text)
   • Select the model to use (base model is built in)
   • Click "Start Recognition"
   • Wait for completion — results will appear in the text box at the bottom of the page

5. GPU acceleration If the machine has an NVIDIA GPU with CUDA properly configured, GPU acceleration will be used automatically.

Linux and macOS Source Code Deployment

Requirements

• Python 3.9 to 3.11
• pip package manager
• FFmpeg (must be installed manually)

Deployment Steps

1. Create project directory and clone the source code

   Create an empty directory (e.g., E:/stt), open a terminal in that directory (on Windows, type cmd in the address bar and press Enter), then run:

   git clone [email protected]:jianchang512/stt.git .

2. Create and activate a virtual environment

   # Create virtual environment
   python -m venv venv
   
   # Activate (choose based on OS)
   # Windows:
   %cd%/venv/scripts/activate
   # Linux/Mac:
   source ./venv/bin/activate

3. Install dependencies

   pip install -r requirements.txt

   Tip: If you encounter version conflicts, try pip install -r requirements.txt --no-deps to skip dependency checks.

4. Install FFmpeg

   • Windows: Extract ffmpeg.7z and place ffmpeg.exe and ffprobe.exe in the project root directory
   • Linux/Mac: Download from the FFmpeg official site and place the ffmpeg and ffprobe binaries in the project root directory

5. Download the model

   Visit the model download page and download the desired model. Place the .pt file in the models folder inside the project root directory.

6. Start the service

   python start.py

   Wait for the browser window to open automatically.

API Reference

STT provides a RESTful API for integration with other tools.

Endpoint

POST {url}/api

Request Parameters

Parameter	Type	Required	Description
file	File	Yes	Audio file (WAV format, 16kHz sample rate, mono)
language	String	No	Language code (e.g., zh, en). Leave empty for auto-detect.
model	String	No	Model name (e.g., base, small)
response_format	String	No	Output format: json, srt, text

Response Format

{
  "code": 0,
  "data": "Recognized text or SRT content"
}

• code is 0 for success, 1 for failure
• data contains the recognition result

Retry Mechanism

The API supports automatic retry — if recognition fails, it will automatically retry.

Common Issues

Issue	Possible Cause	Solution
Browser doesn't open automatically	Firewall block or port in use	Manually visit http://localhost:7860
Empty recognition results	Unsupported audio format	Convert audio to WAV format (16kHz, mono) using FFmpeg
Very slow recognition	GPU acceleration not enabled	Check CUDA environment; ensure NVIDIA drivers and CUDA Toolkit are installed correctly
Insufficient VRAM error	GPU VRAM not enough	Switch to a smaller model (e.g., base or small)
Dependency installation failed	Python version incompatibility	Ensure Python 3.9–3.11 is being used
FFmpeg not found	FFmpeg not placed correctly	Ensure ffmpeg and ffprobe are in the project root directory