VoxFlash-TTS ⚡

July 14, 2026 · View on GitHub

The Fastest Voice Cloning System for Real-Time Inference
Zero-shot · Chinese & English · Edge-Deployable · Consumer GPU

🌐 Live Demo · 🚀 Quick Start · 🏗 Architecture · 📊 Evaluation · 🤗 Model Card

VoxFlash Banner

English · 中文

What is VoxFlash-TTS?

VoxFlash-TTS is the fastest voice cloning system in the industry, built around a radically compressed latent diffusion architecture. It supports zero-shot voice cloning in Chinese and English, runs on consumer-grade GPUs, and is designed from the ground up for edge deployment.

The key insight: most TTS inference bottlenecks are a sequence length problem. By compressing 24kHz audio into a 9 Hz latent representation — roughly 8× more compressed than EnCodec — VoxFlash reduces end-to-end compute by orders of magnitude without sacrificing acceptable audio quality.

Generating 10 seconds of audio requires processing just 90 latent vectors, compared to 750+ in conventional systems.

Highlights

⚡ Millisecond-level inference on consumer GPUs
🎙️ Zero-shot voice cloning — no fine-tuning required
🌏 Chinese & English — same-language and cross-lingual cloning
💻 Edge-friendly — low VRAM footprint, low-end GPU compatible
🐳 One-command Docker deployment — up and running in minutes
📦 ~600 MB ONNX — full pipeline in a single deployable artifact

Architecture

VoxFlash-TTS is built on an ultra-compressed latent diffusion pipeline:

Text Input
    │
    ▼
┌─────────────────────┐
│   Phoneme Encoder   │  ConvNeXtV2 — lightweight, hardware-friendly
└─────────────────────┘
    │
    ▼
┌─────────────────────┐
│  Coarse Alignment   │  Explicit alignment — lower complexity than Cross-Attention
└─────────────────────┘
    │
    ▼
┌─────────────────────┐
│   Diffusion Model   │  Multi-step denoising in latent space (NFE=16)
└─────────────────────┘
    │                     ▲
    │              ┌──────┴──────┐
    │              │Speaker Enc. │  Reference audio → speaker embedding
    │              └─────────────┘
    ▼
┌─────────────────────┐
│    VAE Decoder      │  9 Hz latent → 24kHz high-fidelity waveform
└─────────────────────┘
    │
    ▼
Audio Output

Why 9 Hz?

System	Latent Frame Rate	Latent vectors for 10s audio
EnCodec	~75 fps	~750
Speech LM (semantic tokens)	~50 fps	~500
Stable Audio	~21.5 fps	~215
VoxFlash-TTS	9 fps	90

Transformer self-attention scales at O(n²) with sequence length. Cutting the sequence by 8× reduces attention compute by ~64×. This is why VoxFlash can deliver millisecond inference where others cannot.

Changelog

2026-07-14: Weights updated. See huggingface.co/VoxFlashTTS/VoxFlashTTS

Quick Start

Requirements

NVIDIA GPU with CUDA ≥ 12.3.2
Docker

Installation

# Pull the image
docker pull berlinisaiah/ttsv2:v3

Run

# Background mode (production)
docker container run -d --gpus all \
  --mount type=bind,source=$(pwd)/resources,target=/app/resources \
  -p 8000:8000 berlinisaiah/ttsv2:v3

# Foreground mode (debug)
docker container run -it --gpus all \
  --mount type=bind,source=$(pwd)/resources,target=/app/resources \
  -p 8000:8000 berlinisaiah/ttsv2:v3

Access WebUI

Open your browser and navigate to:

http://127.0.0.1:8000/demo.html

Capabilities

Supported Languages

Language	Same-language Cloning	Cross-lingual Cloning
Chinese (Mandarin)	✅	✅
English	✅	✅

Zero-Shot Cloning

No fine-tuning needed. Provide any reference audio clip and VoxFlash extracts a speaker embedding, injects it into the diffusion process, and outputs speech matching the target voice.

Cross-lingual cloning (e.g. Chinese reference → English output) is supported, demonstrating effective disentanglement of voice timbre from language identity.

Evaluation

Audio samples are drawn from the Seed-TTS evaluation set for direct comparison with leading systems.

System	Inference Speed	Deployment	Zero-Shot	Cross-lingual
Seed-TTS	Slow	Cloud GPU	✅	✅
CosyVoice 2	Medium	Medium	✅	✅
FastSpeech variants	Fast	Low	❌	❌
VoxFlash-TTS	Fastest	Edge / Consumer GPU	✅	✅

👉 Listen to audio samples at voxflash.github.io

Use Cases

Scenario	Key Requirement	VoxFlash Advantage
Real-time voice interaction	Low first-packet latency	Short latent sequences, fewer diffusion steps
Large-scale batch synthesis	Throughput & GPU cost	Orders-of-magnitude compute reduction
Edge / on-device deployment	Low VRAM & power draw	Lightweight architecture, consumer GPU capable
Individual developers	Simple setup	One Docker command, no tuning required

Model Size

File	Size	Contents
`main_model.onnx`	697 MB	Phoneme Encoder + Diffusion Model + Speaker Encoder
`vae_decode.onnx`	51.5 MB	VAE Decoder
`vae_encode.onnx`	46.1 MB	VAE Encoder
`vocoder.onnx`	59.7 MB	Vocoder
Total	~854 MB	Full pipeline

Limitations

Audio quality under extreme compression may fall short of quality-focused systems such as Seed-TTS
Currently optimized for Chinese and English; other languages have not been systematically evaluated
Accent naturalness in cross-lingual cloning has room for improvement
Reference audio shorter than 3 seconds may reduce speaker similarity

Citation

If VoxFlash-TTS has been useful in your research or engineering work, please cite:

@misc{voxflash2026,
  title     = {VoxFlash-TTS: Ultra-Compressed Latent Diffusion for Real-Time Voice Cloning},
  author    = {VoxFlash},
  year      = {2026},
  url       = {https://github.com/VoxFlash/VoxFlashTTS},
  note      = {GitHub repository}
}

📧 Email: zhangtaiyan072@gmail.com
🌐 Demo: voxflash.github.io
🤗 Hugging Face: huggingface.co/VoxFlashTTS/VoxFlashTTS

WeChat Community

Scan the QR code below to join the VoxFlash-TTS Voice Cloning Group on WeChat:

VoxFlash-TTS WeChat Group QR Code

QR code valid for 7 days. Please re-enter the group if it expires.

Not the most expressive TTS — the fastest, lightest, and easiest-to-deploy voice cloning system.