vocalflow 0.7.0

Local TTS and audio transcription web app

VocalFlow

Local TTS, voice cloning, and transcription for Windows.

---

Self-hosted web app that runs entirely on your machine. No cloud APIs, no data leaves your PC.

Features

• Voice Cloning — clone any voice from a short audio clip (3-10s), save & reuse voice prompts
• Custom Voice — 9 preset speakers with emotion/tone control ("say it angrily", "whisper softly")
• Voice Design — create new voices from text descriptions, no reference audio needed
• Transcription — Whisper-powered transcription with word-level timestamps, 6 model sizes
• Smart GPU — automatic model load/unload between switches, only one model in VRAM at a time
• Flash Attention 2 for faster inference
• 10+ languages with auto-detection

Requirements

• Windows 10/11 with a CUDA GPU (6+ GB VRAM)
• Python 3.11
• FFmpeg — winget install ffmpeg
• SoX — winget install sox

Quick Start

pip install vocalflow
vocalflow

Open http://localhost:5001. Models download automatically on first use.

From source

git clone https://github.com/0xBinayak/VocalFlow.git
cd VocalFlow
uv sync
uv run app.py

For auto-reload during development: uv run gradio app.py

Models

Model	Params	Use
Qwen3-TTS-1.7B-Base	1.7B	Voice cloning (best quality)
Qwen3-TTS-0.6B-Base	0.6B	Voice cloning (faster)
Qwen3-TTS-1.7B-CustomVoice	1.7B	Preset speakers + instructions
Qwen3-TTS-0.6B-CustomVoice	0.6B	Preset speakers only
Qwen3-TTS-1.7B-VoiceDesign	1.7B	Voice from text description
Whisper (tiny-turbo)	39M-1.5B	Transcription

All TTS models run in bfloat16 with SDPA/Flash Attention. Whisper falls back to CPU if no GPU.

Contributing

1. Fork & clone, uv sync, create a branch
2. Ensure uvx ruff check app.py transcribe.py main.py passes
3. Open a PR against main

Open an issue if you find a bug.

License

MIT

Acknowledgments

Qwen3-TTS | OpenAI Whisper | Gradio