About Maya 1

Maya 1 is a speech model designed for expressive voice generation with human emotion and precise voice design. Developed by Maya Research, this open source text-to-speech system enables natural voice generation with over 20 emotions, real-time streaming capabilities, and natural language voice control.

What is Maya 1?

Maya 1 is a 3 billion parameter text-to-speech model built on a Llama-style transformer architecture. Unlike traditional TTS systems that produce flat, robotic voices, Maya 1 generates speech with emotional depth and personality. The model uses the SNAC neural codec to produce high-quality 24 kHz audio at approximately 0.98 kbps, making it suitable for real-time streaming applications.

The model's key innovation is its ability to understand natural language voice descriptions. Instead of adjusting complex technical parameters, users describe voices in everyday language, such as "40-year-old, warm, low pitch, conversational" or "Female, in her 30s with an American accent, energetic, clear diction." The model interprets these descriptions and generates voices that match the specifications.

Key Features

• Natural Language Voice Control: Describe voices using everyday language, just like briefing a voice actor
• 20+ Emotions: Support for laugh, cry, whisper, angry, sigh, gasp, giggle, chuckle, and many more emotional expressions
• Real-Time Streaming: SNAC neural codec enables real-time voice synthesis with low latency
• Production-Ready: Runs on single GPU, integrates with vLLM for scalable deployments
• Open Source: Apache 2.0 license allows commercial use, modification, and distribution
• Multi-Accent Support: English with various accents including American, British, and Middle Eastern
• Customizable: Architecture supports fine-tuning on custom datasets for specialized voices

Technical Architecture

Maya 1 uses a 3 billion parameter decoder-only transformer architecture, similar to Llama models, but trained to predict SNAC neural codec tokens instead of text tokens. The generation process follows this pipeline: voice description and text are tokenized, the model generates SNAC audio code tokens (7 tokens per frame), and these tokens are decoded into 24 kHz audio.

The SNAC neural codec uses a multi-scale hierarchical structure operating at approximately 12, 23, and 47 Hz. This keeps autoregressive sequences compact, enabling real-time streaming at about 0.98 kbps while maintaining high audio quality.

Training and Development

Training involved two main phases. First, pretraining on an internet-scale English speech corpus provided broad acoustic coverage and natural coarticulation patterns. Second, supervised fine-tuning on a curated dataset of studio recordings included human-verified voice descriptions, 20+ emotion tags per sample, multi-accent English coverage, and character and role variations.

The data pipeline ensured quality through 24 kHz mono resampling with -23 LUFS normalization, voice activity detection with silence trimming, forced alignment for clean phrase boundaries, and both text and audio deduplication.

Mission and Vision

Maya Research builds emotionally intelligent, native voice models that enable the rest of the world to speak. The organization believes voice intelligence should not be a privilege reserved for the few. Current voice models only work well for a narrow slice of English speakers because training data for most accents, languages, and speaking styles simply doesn't exist.

By making Maya 1 open source, Maya Research aims to accelerate research and enable developers worldwide to build on this work. The goal is to create voice intelligence for everyone, especially the 90% of the world ignored by mainstream voice AI systems.