Clone ANY Voice for Free — Qwen Just Changed Everything

Open-source voice cloning and “voice design” have moved from closed, API-only systems into the open TTS ecosystem: Qwen has released its Quen 3 TTS family as open weights, including tools for cloning, designing, and generating voices. That shift matters because it lets developers download models, run them locally, and build multilingual voice experiences without being locked into a single vendor’s interface—an important change after months of rapid TTS progress that largely stayed behind paywalled access.

The Quen 3 TTS family arrives with two model sizes. A smaller 0.6B model supports multiple languages and streaming, and—crucially—comes with both a base model and a fine-tune model. That pairing is positioned as a practical on-ramp for creating custom voices: instead of only using pre-made speaker presets, developers can fine-tune from the open base. The larger 1.7B model retains the smaller model’s multilingual capabilities while adding instruction control for voice design and voice cloning. In practice, that means text plus an instruction prompt can shape the output voice (for example, describing a “young anime” tone or a “documentary” delivery), and voice cloning can be triggered by providing a short audio sample so the system extracts a voice representation and renders new speech in that voice.

A major theme is reducing the barrier for non-English users. The 0.6B model supports 10 languages, nine dialects, and 49 “tambas” (as described in the announcements), and the open release includes tokenizers and related components. That openness is framed as enabling training and adaptation for additional languages and dialects—especially relevant when many earlier systems performed best in English.

The transcript also highlights quality and usability features that go beyond “sound it out phonetically.” Because the system is trained on large text corpora and uses a tokenizer/codebook approach, it can pronounce symbols like email addresses and other structured text without requiring manual phonetic transcription. The system is also described as fully end-to-end, with multiple token types (text tokens, codebook tokens, and speaker embeddings) feeding a streaming decoder—contrasting with older pipelines that stitched together separate modules.

Hands-on demos on Hugging Face show the workflow. With the 0.6B model, the user selects pre-made speakers (including Chinese dialect variants) and generates speech in other languages, sometimes with noticeable artifacts (Spanish is called out). The demos include batch inference for faster audio generation and long-form text handling such as dates and numbers. Auto language guessing is also demonstrated, supporting code-switching scenarios where a mostly single-language script can include brief shifts.

With the 1.7B model, instruction prompts drive voice style changes—cartoon, villain, documentary, and emotional variants like neutral, excited, whispering, loud/soft, and dramatic. Voice cloning is demonstrated using roughly a 10-second sample audio, producing new speech in the cloned voice. The results are described as impressive but not perfect, with the transcript noting that real-world audio cleanup (e.g., noise reduction) can affect the clone’s fidelity.

Finally, the release includes a Hugging Face space for trying the models and a paper referencing training on over 5 million hours of speech data. The overall takeaway is that Qwen’s open-sourced Quen 3 TTS family makes advanced voice cloning and instruction-driven voice generation more accessible for experimentation, fine-tuning, and future deployment on smaller or edge-focused runtimes.