Maya Research

Maya: A Native AI for Speech

Maya Research develops native AI speech models designed for the languages that most models get wrong. Maya builds real-time speech technology inside the language rather than translating into it, preserving natural accent, rhythm, and stress.

Key Features

• Native Speech Generation: Delivery follows the intent (warm, urgent, or apologetic) instead of flattening to one register. Supports over 10 languages including Hindi, Telugu, Arabic, Tamil, English, Bengali, Indonesian, Marathi, Kannada, Malayalam, Gujarati, Punjabi, and Odia.

• Low Latency Performance: Features a 75ms time to first byte (p50) with a 7.3ms standard deviation. Audio starts before generation finishes, meaning a paragraph begins speaking as fast as a sentence.

• Maya 2 Native: The flagship model.

• Maya 2 Global: Offers wider coverage for teams shipping beyond one market.

• Maya 2 Flash: The speed tier offering the lowest latency of the family.

• Open Weights Models: Includes Maya 1 (Apache 2.0 license) and Veena (an earlier Hindi and English model) for users to run, fine-tune, and ship in their own stack.

• Streamlined API: A single endpoint API with no sessions to open or jobs to poll, outputting 24 kHz streamed PCM. Hosted in Mumbai and Iowa regions.

Use Cases

• Conversational AI: Customer support, voice agents, and IVR systems.

• Industry Applications: BFSI and E-commerce.

• Media: Storytelling and announcements.

Getting Started

Website: https://www.mayaresearch.ai/

Maya provides speech models built from the land of languages, offering both hosted API solutions and open-weight models for developers.