STT — Voice Recognition to Text Tool

// Overview

STT is an offline speech-recognition tool that turns the human voice in an audio or video file into text. It is built on the open-source faster-whisper models and outputs three formats: JSON, SRT subtitles with timestamps, and plain text. The whole pipeline runs on the local machine — nothing is uploaded — so it can be deployed on an intranet or an air-gapped host.

The interface is a local web page rather than a cloud service: you start the program, it opens a browser window, and you drag an audio or video file onto the upload area, choose the spoken language, the output format and the model, and start recognition. Results appear in the page in the format you picked. A precompiled Windows build ships as a double-clickable start.exe; Linux and macOS run from source.

The project positions itself as a self-deployed replacement for a hosted speech-recognition API — the README's claim is that accuracy is essentially the same as OpenAI's own endpoint, because it is the same Whisper model family running locally. Model size is a deliberate trade: tiny, base, small, medium and large-v3 are all supported, quality climbs with size and so does the hardware requirement, and one of the smallest models ships with the program so it works before you download anything — the larger ones are separate downloads you unpack into the models directory. On a machine with an NVIDIA GPU and a correctly configured CUDA environment, CUDA acceleration is used automatically. It comes from the author of pyVideoTrans and shares that project's community.

// What it does

• Fully offline: recognition runs locally, with no network round-trip, suitable for intranet or air-gapped deployment
• Three output formats — JSON, timestamped SRT subtitles and plain text
• The full faster-whisper model ladder (tiny, base, small, medium, large-v3), with one of the smallest bundled so it works out of the box
• Automatic CUDA acceleration when an NVIDIA GPU and CUDA environment are present
• A local browser UI with drag-and-drop upload, language selection and format selection
• Precompiled Windows build (start.exe) alongside source deployment for Linux and macOS

// Getting started

On Windows the quickest path is the precompiled release. On Linux and macOS you deploy from source with Python 3.9–3.11; either way the tool ends up serving a local page in your browser.

https://github.com/jianchang512/stt/releases

git clone https://github.com/jianchang512/stt.git .
python -m venv venv
source ./venv/bin/activate
pip install -r requirements.txt

pip uninstall -y torch
pip install torch --index-url https://download.pytorch.org/whl/cu121

python start.py

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Generate timestamped SRT subtitles for a video archive without sending the media to a third party
• Stand up an internal transcription endpoint on an intranet where cloud speech APIs are not permitted
• Trade accuracy against hardware by picking a model size, from the bundled starter model up to large-v3
• Transcribe interviews or recordings on a laptop, using CUDA when a GPU is available and CPU when it is not

// How STT — Voice Recognition to Text Tool compares

STT — Voice Recognition to Text Tool alongside other open-source audio, music & voice tools AI/TLDR tracks, ranked by GitHub stars.