# Clone ANY Voice for Free — Qwen Just Changed Everything

Source: https://www.youtube.com/watch?v=jZ8wPB-KI8g
Recap page: https://rapidrecap.app/video/jZ8wPB-KI8g
Generated: 2026-01-23T16:06:24.953+00:00

---
## Quick Overview

Qwen-TTS models, including the newly open-sourced Qwen3-TTS family featuring VoiceDesign, VoiceClone, and Generation capabilities, are demonstrated to achieve state-of-the-art performance across multilingual TTS tasks, offering fine-grained control over voice characteristics and supporting 10 languages and 9 dialects.

**Key Points:**
- Qwen3-TTS models are now fully open-sourced on GitHub and accessible via the Qwen API, supporting 10 main languages and 9 dialects.
- The Qwen3-TTS-12Hz-1.7B-VoiceDesign model allows users to describe voices using natural language instructions, creating entirely new voices.
- The Qwen3-TTS-12Hz-1.7B-Base model supports 3-second rapid voice cloning from reference audio.
- The CustomVoice model (1.7B and 0.6B variants) adds instruction control for emotion, tone, and speaking style (e.g., happy, sad, whispering, loud) on preset voices.
- The Qwen-TTS-Tokenizer-12Hz achieves state-of-the-art performance in speech reconstruction metrics like PESQ (3.21 wideband) and STOI (0.96), outperforming similar tokenizers.
- The demonstration includes generating multilingual speech, comparing different speaking styles, and showcasing the Voice Design workflow using a custom 'Wise Old Wizard' voice.
- The model architecture utilizes a dual-track LM architecture with two specialized tokenizers (Qwen3-TTS-Tokenizer-25Hz and Qwen-TTS-Tokenizer-12Hz) for efficient streaming generation.

![Screenshot at 00:13: The initial slide introduces the Qwen3-TTS models' capabilities, highlighting VoiceDesign, VoiceCloning, and Generation, setting the stage for the subsequent technical demonstrations.](https://ss.rapidrecap.app/screens/jZ8wPB-KI8g/00-00-13.jpg)

**Context:** The video serves as a technical showcase and tutorial for the Qwen3-TTS family of text-to-speech models, recently released open-source by the Qwen Team. The presentation highlights the models' capabilities in multilingual synthesis, voice cloning, and fine-grained voice design via natural language instructions. It reviews the technical report, compares performance benchmarks against other leading models, and provides runnable code examples demonstrating basic TTS, multilingual output, emotion/style control, voice cloning from reference audio, and the reusable voice prompt workflow.

## Detailed Analysis

The video details the capabilities of the newly open-sourced Qwen3-TTS family of models, emphasizing their multilingual, controllable, robust, and streaming text-to-speech features. The speaker reviews technical achievements, noting that the models support 10 main languages and 9 dialects. Key models include the 1.7B and 0.6B variants, featuring VoiceDesign for creating custom voices via natural language descriptions (e.g., 'Wise Old Wizard' voice), Instruction Control (CustomVoice model) for altering emotion/style on preset voices, and Base models for 3-second voice cloning. Benchmarks show the Qwen3-TTS-Tokenizer achieving SOTA results on the LibriSpeech test set, outperforming competitors in PESQ and STOI scores. The demonstration walks through code examples showing basic TTS, multilingual synthesis using the Ryan speaker, emotion control using the Serena speaker, custom voice design producing the 'Wise Old Wizard' voice, and voice cloning using a 3-second reference audio sample. The presenter concludes by noting the impressive nature of these advancements, which were previously restricted to models from major labs like OpenAI and Google, and that the open-sourcing of these models marks a significant step forward for the community.

### Initial Announcements

- Qwen3-TTS family announced VoiceDesign and VoiceClone in June; Qwen3-TTS-Flash released in September supporting multi-timbre, multilingual, multi-dialect synthesis
- Qwen3-TTS Family Open Sourced announced in January 2026

### Model Capabilities Overview

- Qwen3-TTS supports 10 languages, 9 premium voices, instruction control (1.7B model), streaming generation, and high-quality ASR-resistant performance

### Model List Overview

- Detailed table shows 1.7B and 0.6B models supporting Streaming, Multilinguality, Voice Clone, and Instruction Following features

### Tokenizer Performance

- Qwen-TTS-Tokenizer-12Hz achieved SOTA scores: PESQ_WB 3.21, PESQ_NB 3.68, STOI 0.96, UTMOS 4.16, SIM 0.95, surpassing competitors like SpeechTokenizer and Mimi

### Basic TTS Generation

- Demonstrates generating speech in English (Ryan speaker) and Chinese (Vivian speaker) using default settings

### Multi-Speaker Demo

- Shows Ryan speaker generating text in English, German, French, Spanish, and Italian, and native speakers (Ono_Anna, Sohee, Dylan, Eric) generating their native languages

### Voice Design Model

- Demonstrates creating custom voices via natural language instructions, generating a 'Cute Anime Girl Voice' and a 'Deep Authoritative Narrator' voice (David Attenborough style) using the 1.7B VoiceDesign model

### Instruction Control with CustomVoice

- Shows Ryan speaker modifying emotion (Happy, Sad, Angry, Fearful, Neutral) using the 1.7B CustomVoice model on the same text input ('I just found out the news about what happened yesterday.') to produce different emotional outputs, including a successful neutral output demonstrating baseline quality control. The custom voice cloning workflow is also introduced, allowing users to clone a voice from a reference audio file and apply it to new text, as demonstrated by cloning the 'Wise Old Wizard' voice and using it to generate dialogue, and then demonstrating batch generation using the cloned voice prompt for efficiency.

![Screenshot at 00:00: Illustration depicting Text-to-Speech \(TTS\) models exploding from a volcano, symbolizing a major, powerful release in the TTS field.](https://ss.rapidrecap.app/screens/jZ8wPB-KI8g/00-00-00.jpg)
![Screenshot at 00:14: A summary slide listing key features of the Qwen3-TTS models, including VoiceDesign, Instruction Control, Voice Cloning, and performance metrics.](https://ss.rapidrecap.app/screens/jZ8wPB-KI8g/00-00-14.jpg)
![Screenshot at 07:24: A comparison table showing Qwen3-TTS-12Hz-1.7B achieving superior or competitive scores across various benchmarks for Voice Clone, Voice Design, and Custom Voice tasks against other models like Qwen3 Omni and MiniMax.](https://ss.rapidrecap.app/screens/jZ8wPB-KI8g/00-07-24.jpg)
![Screenshot at 08:00: Hugging Face repository interface showing the Qwen3-TTS collection with various models available, including CustomVoice, Base, and VoiceDesign variants.](https://ss.rapidrecap.app/screens/jZ8wPB-KI8g/00-08-00.jpg)
![Screenshot at 11:20: Code demonstrating Long-Form Text Generation using the Serena speaker to synthesize a short story excerpt, highlighting the model's ability to maintain coherence over longer passages.](https://ss.rapidrecap.app/screens/jZ8wPB-KI8g/00-11-20.jpg)