Qwen3-TTS: The ElevenLabs Killer?

Quint 3 Text-to-Speech System: A Detailed Overview

Key Concepts:

• Quinn 3: The latest open-weight text-to-speech (TTS) system released by the Quint team.
• Text-to-Speech (TTS): Converting written text into spoken audio.
• Voice Cloning: Replicating a specific voice based on a short audio sample.
• Voice Design: Creating custom voices based on textual descriptions.
• Open-Weight Models: Machine learning models with publicly available weights, allowing for customization and research.
• Tokenizer (Quenfree TTS Tokenizer 12Hz): A component that breaks down text into smaller units for processing by the TTS model, enabling low-latency streaming audio.
• VRAM: Video Random Access Memory, the memory on a graphics card used for processing.
• MLX: A machine learning framework optimized for Apple Silicon.

1. Introduction & Capabilities

The Quint team has released Quinn 3, a new open-weight TTS system boasting three core capabilities: text-to-speech generation, voice cloning, and voice design. The speaker highlights Quinn 3’s voice cloning as being superior to other open-weight models tested. A key feature demonstrated is the system’s ability to generate coherent conversations between multiple characters, maintaining voice consistency throughout extended dialogues. An example conversation between two characters was presented, showcasing this capability.

2. Model Families & Language Support

Quinn 3 is available in two model families: a 1.7 billion parameter model and a 6 billion parameter model. The release is geared towards enabling deployment on edge devices. Currently, the system supports 10 different languages. The 1.7 billion model is more customizable, offering base versions, custom voice creation options, and voice design capabilities. Notably, Quinn 3 is the first open-weight model allowing control over output behavior and tone using text prompts, similar to the Gemini speech generation API. The system surpasses previous open-weight models on key benchmarks, though detailed benchmark results are linked in the video description.

3. Technical Foundation & Tokenization

The development of Quinn 3 involved building a custom tokenizer, the “Quenfree TTS Tokenizer 12Hz.” This tokenizer facilitates low-latency streaming audio generation controllable via text prompts. The speaker mentioned potential for a follow-up video detailing the technical aspects further.

4. Practical Implementation & Notebook Walkthrough

A Colab notebook was created to demonstrate the system’s features. The notebook supports 10 languages and is designed for ultra-low latency. It allows for voice cloning, custom voice design using natural language descriptions, and utilizes both the 6B and 1.7B models, runnable on a free Google Colab instance. An MLX version is also available for Mac OS users, created by Prince Kuma. Running the notebook requires selecting a T4 GPU runtime.

5. Text-to-Speech Generation & Emotion Control

The notebook begins with standard speech generation, offering nine pre-defined voices. While optimized for native speaker languages, the system can generate audio in other languages as well. Emotion control is achieved by passing a dictionary containing emotions and their descriptions. The example demonstrated generating speech with neutral, happy, angry, and sad emotions. The speaker noted occasional “drift” in output quality, emphasizing the probabilistic nature of the system and the need for multiple runs to achieve desired results.

6. Custom Voice Design

The notebook then showcases voice design, utilizing the 1.7B model, which is the only model in the current release that effectively follows instructions. The process involves providing detailed text descriptions of the desired voice characteristics. Examples included creating voices for a “wise old wizard,” an “energetic anime girl,” and a “news anchor.” The generated outputs were presented, with the “anime girl” example being particularly impressive, even in Japanese (despite the speaker’s lack of Japanese proficiency).

7. Voice Cloning Capabilities

Quinn 3’s voice cloning functionality was demonstrated, claiming high-quality clones can be generated from as little as 3 seconds of audio. The example used a 9-second audio sample of the speaker’s voice. Three audio segments were generated using the cloned voice, with the speaker noting the third segment (“I can’t believe how realistic this voice cloning technology has become.”) was the most accurate. The speaker requested viewers to compare the cloned audio to their original voice and provide feedback in the comments.

8. Combined Voice Design & Cloning for Narration

The final section of the notebook demonstrated combining voice design and cloning for narration. A character description was used to generate a reference audio, which was then used to create a clone prompt for narration. A short dialogue was generated using the cloned voice, exhibiting improved consistency compared to previous cloning attempts. The speaker highlighted the system’s ability to generate streaming responses, potentially enabling real-time interactions (though this functionality wasn’t implemented in the notebook).

9. Performance & Resource Requirements

The non-streaming version of the system takes approximately 10-15 seconds to generate 7 seconds of audio. GPU RAM usage can be high (around 3-4 GB VRAM for a single model loaded), but can be reduced by loading only one model at a time. The MLX version allows for usage on Mac OS devices.

10. Conclusion & Resources

Quinn 3 represents a significant advancement in open-weight TTS technology, offering powerful capabilities in text-to-speech generation, voice cloning, and voice design. The speaker encouraged viewers to explore the Colab notebook (link in the video description) and highlighted its potential for applications like voice agents (referencing their project Verby, with over 1000 stars on GitHub) and local speech dictation systems (their project Write). The speaker expressed interest in creating a follow-up video detailing the technical aspects of the system and encouraged viewers to indicate their interest in streaming audio output functionality.