SWT (声文通) Local Offline Speech-to-Text Desktop Tool — Secure, Fast, No Data Upload

ydtg1993 — Thu, 09 Jul 2026 05:14:53 +0000

SWT (声文通)

SWT is a fully offline speech-to-text desktop application built on the FunASR speech recognition framework. All audio data is processed locally — no internet connection required, and nothing is uploaded to any server.

Download

Preview

Features

Feature	Description
File Transcription	Import audio (MP3 / WAV / FLAC / …) or video (MP4 / MKV / …) files, auto-extract audio and output text
Batch Processing	Multi-file sequential transcription, stable and reliable
Multi-Model Support	6 built-in ASR models, one-click download & switch, CPU/GPU inference supported
Multilingual	Supports Chinese, English, Japanese, Korean, Cantonese, and more
Result Export	Export to TXT / JSON / SRT formats, ready for subtitles
Speaker Diarization	Experimental multi-speaker recognition (based on CAM++)
VAD Voice Activity Detection	FSMN-VAD intelligent long audio segmentation, adjustable silence threshold and segment length
Speech Enhancement	AI denoising (FRCRN / DFSMN), echo cancellation (DFSMN-AEC), speech separation (MossFormer)
Post-Processing	CT-Transformer punctuation restoration, LLM AI correction of typos and grammar issues

Built-in Models

Model	Size	GPU Recommended	CPU	Description
Qwen3-ASR (0.6B)	~1.8 GB	2GB+	Smooth	Lightweight multilingual, entry-level
Qwen3-ASR (1.7B)	~3.5 GB	6GB+	Runs but slow	High-accuracy Chinese recognition, built-in punctuation
SenseVoiceSmall	~0.5 GB	Not required	Smooth	Ultra-lightweight, multilingual + emotion/event detection
Whisper-large-v3	~3.1 GB	8GB+	Not recommended	OpenAI flagship multilingual model
faster-whisper-tiny	~0.2 GB	Not required	Smooth	CTranslate2 accelerated, extremely lightweight
Fun-ASR-Nano	~1.0 GB	2GB+	Runnable	FunAudioLLM lightweight ASR, Chinese optimized

System Requirements

Item	Requirement
OS	Windows 10 / Windows 11
Runtime	Python 3.10+
Hardware	CPU-capable; NVIDIA GPU acceleration supported (CUDA 12.1+)
Storage	~0.2–3.5 GB per model; 10 GB free disk space recommended
Dependencies	ffmpeg (required for video transcription; optional for audio-only)

Quick Start

1. Clone and Enter Project

git clone https://gitee.com/ydtg1993/swt.git
cd swt

2. Create Virtual Environment and Install Dependencies

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

3. Download Models

At least one ASR model must be downloaded on first run. After launching the application, go to Settings → Model Download Manager and select the desired models to download.

Models will be saved in the llm/ directory.

4. Launch Application

python main.py

Note: Transcribing video files requires installing ffmpeg and adding it to your system PATH.

Project Structure

swt/
├── main.py                  # Application entry point
├── requirements.txt         # Python dependencies
├── LICENSE                  # MIT License
├── config/                  # Configuration (YAML + Pydantic Settings)
├── core/                    # Core engine (ASR, VAD, subtitle generation)
├── models/                  # Data models (SQLAlchemy)
├── ui/                      # GUI layer (PySide6 + QFluentWidgets)
│   ├── pages/               # Page components
│   └── widgets/             # Reusable components
├── workers/                 # Background worker threads (QRunnable)
├── utils/                   # Utility functions
├── llm/                     # Model storage directory
├── resources/               # Icons, images
├── scripts/                 # Model download helper scripts
├── tests/                   # Tests
└── logs/                    # Logs

Tech Stack

Category	Technology
GUI	PySide6 + QFluentWidgets
ASR Engine	FunASR (Qwen3-ASR, SenseVoiceSmall, FSMN-VAD)
Database	SQLite + SQLAlchemy 2.x
Configuration	YAML + QSettings
Logging	loguru
Task System	QThreadPool + QRunnable + Signal
File Handling	pathlib
Packaging	PyInstaller

Development

# Install dependencies
pip install -r requirements.txt

# Run application
python main.py

# Run tests
pytest tests/ -v

# Package as executable
pyinstaller main.spec

Third-Party Licenses

Project	License
FunASR	MIT
Qwen3-ASR	Apache 2.0
SenseVoice	Apache 2.0
faster-whisper	MIT
PySide6	LGPL
QFluentWidgets	GPL / Commercial
ffmpeg	LGPL / GPL
ModelScope	Apache 2.0

License

SWT (声文通) is open-sourced under the MIT License.

Copyright © 2026 SWT (声文通)

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files, to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software.

Project Links

Gitee: https://gitee.com/ydtg1993/swt

Made with ❤️ by ydtg1993

I turned Telegram into a desktop cloud drive — free, unlimited, open source

ydtg1993 — Thu, 18 Jun 2026 16:12:06 +0000

Over the past few weeks I built CraftCloud — a desktop app that turns Telegram into a fully-featured cloud drive. Think Google Drive, but free and unlimited, running on Telegram's infrastructure.

What it does

Full cloud-drive UI — folder tree, list/icon views, drag-and-drop
Upload files up to 4GB each, no total storage limit
Online preview for images, audio (with seek bar), video, text
Auto-sync local folders to Telegram directories on a schedule
Fuzzy search with Whoosh full-text indexing + jieba Chinese word segmentation
6 languages, system tray, dark/light themes
Task queue with persistent history (SQLite + SQLAlchemy ORM)

Tech stack

Layer	Tech
GUI	PySide6 ≥6.5 + PySide6-Fluent-Widgets
Telegram API	Telethon ≥1.36
Database	SQLAlchemy 2.x ORM + SQLite (WAL mode)
Config	Pydantic ≥2.0 + PyYAML
Logging	loguru
Search	Whoosh + jieba
Caching	diskcache

What my project does

Over the past few weeks I built CraftCloud — a desktop app that turns Telegram into a fully-featured cloud drive. Think Google Drive, but free and (virtually) unlimited, running on Telegram's infrastructure. It gives you a native desktop interface to upload, download, search, and manage files directly through your Telegram account, with automatic sync and full local caching.

Target audience

This project is aimed at two main groups:

Telegram power users who want unlimited, zero-cost cloud storage without monthly subscriptions (perfect for backing up photos, videos, and documents).
Python developers / hobbyists looking for a real-world reference on building a complex async GUI desktop app with Telethon, PySide6, and advanced thread/event-loop concurrency control.

It's currently a side-project / production-ready-for-personal-use tool (I use it daily), but it's stable enough for anyone comfortable with self-hosting their own cloud solution on Telegram.

Comparative analysis (vs. existing alternatives)

Alternative	How CraftCloud differs
Google Drive / Dropbox / iCloud	❌ No recurring subscription fees. Telegram offers 2GB per file and effectively unlimited total storage (based on your account). Data is transmitted via Telegram's MTProto encryption.
Rclone (Telegram backend)	✅ CraftCloud provides a native, polished desktop GUI (Windows 11 fluent design) instead of a CLI. Drag-and-drop, visual file management, and built-in search make it accessible to non-developers.
Telegram Desktop native storage	✅ Dedicated file manager view with advanced features like full-text search (Whoosh + jieba), WAL-mode SQLite for reliability, and automatic background sync.

Key trade-offs to consider:

⚠️ Speed is limited by Telegram's datacenter rate limits (not ideal for massive 100GB+ daily transfers).
⚠️ It relies on Telegram's server-side encryption (MTProto), not client-side zero-knowledge E2EE—so treat it as a convenient backup tool rather than a secure vault for highly sensitive data.

DEV Community: ydtg1993