Maya1

Voice design in Maya1 happens in prose: describe a speaker as a raspy 40-year-old baritone or a bright teenage narrator and the 3B model synthesizes that voice directly, no reference audio or cloning session required. Built as a Llama-style transformer, it predicts tokens for the SNAC neural codec, seven per audio frame at about 0.98 kbps, which makes low-latency streaming synthesis natural, and inline tags such as laugh and cry inject emotion mid-sentence for game characters, podcasts, and voice assistants. Maya Research released the weights, tokenizer, and inference scripts on Hugging Face under Apache-2.0 with commercial use allowed, plus a vLLM script for real-time streaming deployments. English is the focus at launch with multi-accent coverage and broader language support on the roadmap. A single GPU with 16GB of VRAM, anything from an RTX 4090 upward, runs it comfortably, and the November 2025 release quickly became the reference point for open description-driven voice generation.