Zyphra has announced the release of ZONOS3-Mini, a compact, open-source text-to-speech (TTS) model designed for high-fidelity voice cloning and natural speech synthesis. The model utilizes a Mixture-of-Experts (MoE) architecture, featuring 800 million total parameters with approximately 100 million active parameters per token.
Technical Specifications and Capabilities
According to technical documentation provided by Zyphra, the model is engineered to run in real time on a variety of hardware, including laptops, mobile phones, and directly within web browsers via WebGPU. The architecture employs a 11-layer backbone with a model dimension of 1,024 and a sequence mixer consisting of nine Gated DeltaNet (GDN) layers and two global attention layers.
ZONOS3-Mini supports 43 languages across several global regions, including Western and Southern Europe, East Asia, Southeast Asia, South Asia, the Middle East, and various parts of Europe. The system uses a SuperBPE tokenizer with 8,192 entries and NeMo normalization for specific languages such as English, Mandarin, Japanese, Korean, German, Spanish, French, and Italian.
The model offers granular control over vocal output, allowing users to adjust emotion, pace, and delivery at the individual word level. Users can also shape character profiles by adjusting age, accent, and vocal texture. The system supports speaker cloning via audio samples of up to 60 seconds or through 192-dimensional MECT-B2 speaker embeddings.
Deployment and Accessibility
Zyphra has released the model weights under the Apache 2.0 license, facilitating open-source integration. The company provides several native runtimes through its GitHub repository, including a CUDA server for NVIDIA GPUs, an MLX server for Apple Silicon, a C++/ggml server for CPU/Metal/Vulkan, and a WebGPU build for browsers.
For users preferring managed services, a hosted version is available through Zyphra Cloud. The model is capable of generating up to 10 minutes of audio per generation at a 44.1 kHz sample rate.
Initial Reception and Performance
Early feedback from the developer community has been mixed regarding the model’s performance relative to established competitors. On the social media platform Reddit, some users noted that while the audio quality in Zyphra’s provided samples appeared high for the model’s size, it did not appear to surpass the quality of existing services such as ElevenLabs or Fish Audio. One user noted that the quality claims made in Zyphra’s blog post appeared inconsistent with the provided benchmarks.
Additionally, some early testers reported technical difficulties when attempting to use the emotion control features, noting that such adjustments could occasionally cause the voice output to break. However, other users on the platform praised the release, specifically citing the utility of the Apache 2.0 licensing.
