OmniVoice: Voice Cloning Across 600 Languages

OmniVoice is an open-source text-to-speech model that clones any voice from a 10-second sample, supports over 600 languages, and produces speech with genuine emotion.

What it is

OmniVoice is an open-source text-to-speech system from the k2-fsa team that focuses on three things: zero-shot voice cloning, broad language coverage, and emotionally expressive output. It is built on the Qwen3-0.6B model and claims support for over 600 languages, with a specific fine-tuned variant for Vietnamese.

The headline feature is zero-shot voice cloning. You provide a 10-second audio sample of any voice, and OmniVoice can generate new speech in that voice without additional training. The model also aims to go beyond flat, robotic TTS output by understanding context and producing speech with emotion, including laughter, sighs, and natural conversational pacing. It runs on NVIDIA GPUs and Apple Silicon Macs, requiring only 16GB of RAM.

Why it matters

• 10-second zero-shot cloning is a low barrier. Many voice cloning systems require minutes of clean reference audio or fine-tuning runs. OmniVoice's 10-second threshold makes cloning accessible for quick experiments and real-world use.
• 600 languages is extraordinary coverage. Most TTS models handle a handful of major languages well. Supporting 600 languages, even with varying quality, makes this relevant for communities and applications that mainstream TTS ignores.
• Emotion-aware output. Traditional TTS produces flat, soulless speech. OmniVoice attempts to understand context and generate output with appropriate emotional cues, including laughter and breathing. If this works as advertised, it narrows the gap between AI speech and human speech significantly.
• Vietnamese-specific fine-tuned model. For Vietnamese developers and content creators, having a dedicated fine-tuned model trained on over 1,000 hours of Vietnamese speech data is significant. The project claims it outperforms existing Vietnamese TTS tools like VieneTTS and Vbee in pronunciation accuracy and expression.
• Runs on consumer hardware. 16GB of RAM and an NVIDIA GPU or Apple Silicon Mac is within reach for most developers. You do not need a data center to run this.

How it works

OmniVoice is built on the Qwen3-0.6B model, a compact language model adapted for speech synthesis. The zero-shot cloning mechanism works by encoding the speaker characteristics from the reference audio sample and conditioning the generation process on that encoding. This means the model captures the voice identity from 10 seconds of audio and applies it to new text without retraining.

The emotion-aware output comes from the model's training on data that includes contextual emotional cues. Rather than producing phonetically correct but emotionally flat speech, the model attempts to match the emotional tone of the input text, generating laughter, pauses, and emphasis where appropriate.