Human-1: The First Full-Duplex Conversational AI System for Hindi

A Full-Duplex Conversation Breakthrough for Indian Languages

In the field of voice conversational AI, full-duplex technology allows systems to "listen and speak simultaneously" like a real person, supporting natural interactive behaviors such as interruptions, overlapping speech, and backchannel responses. However, this cutting-edge technology has long been concentrated almost exclusively on mainstream languages like English, leaving Indian languages virtually unserved.

Recently, the research team at Josh Talks, a well-known Indian content platform, published a paper on arXiv (arXiv:2604.23295) officially introducing a full-duplex conversational modeling framework called "Human-1." It is the first open, reproducible full-duplex voice conversation system designed for Hindi, marking an important milestone for low-resource languages in conversational AI.

Core Technology: Hindi Adaptation Based on the Moshi Architecture

Human-1's technical foundation is derived from Moshi — one of the most advanced full-duplex voice conversation architectures available today. Moshi was released by French AI laboratory Kyutai in 2024 and is capable of simultaneously processing both input and output audio streams within a single model, enabling truly real-time bidirectional conversation.

To migrate this architecture to Hindi, the research team carried out several key adaptations:

• Custom Hindi Tokenizer: A specialized tokenization scheme was designed around Hindi's phonological and orthographic characteristics to better capture the language's phonetic units and semantic structures.
• Large-Scale Real Conversational Data Training: The team collected approximately 26,000 hours of real spontaneous conversational data for model training. Sourced from real-world conversations accumulated on the Josh Talks platform, the data encompasses rich phenomena found in natural spoken language — including interruptions, overlapping speech, and discourse markers — far surpassing broadcast-quality read speech data.
• Open and Reproducible Design: The team emphasized the system's openness and reproducibility, providing a solid foundation for future researchers to extend the work to other Indian languages.

Why Full-Duplex Conversation Matters

Current mainstream voice conversation systems, including most voice assistants built on large language models, are essentially "half-duplex" — the system must wait for the user to finish speaking before responding and cannot provide real-time feedback while the user is talking. This creates a significant gap from the real human conversational experience.

The core value of full-duplex technology lies in:

1. Enhanced Naturalness: Supports real-time insertion of backchannel cues like "uh-huh" and "right," making conversations more fluid and natural.
2. Interruption Handling: Users can interrupt the system at any time, and the system responds instantly.
3. Overlapping Speech: Allows both parties to speak simultaneously, closely mirroring real-world communication.

For a market like India with over 600 million Hindi speakers, the deployment of full-duplex conversation technology will directly impact user experience across numerous application scenarios, including voice assistants, customer service bots, and educational companions.

Industry Significance and Future Outlook

The release of Human-1 carries multiple layers of significance. First, it demonstrates that cutting-edge voice conversation technology can be effectively transferred to low-resource languages, offering a technical roadmap for dozens of "neglected" languages worldwide. Second, Josh Talks' foray into foundational AI research as a content platform highlights the unique advantage that companies with real data resources hold in the AI era — 26,000 hours of authentic conversational data is an extremely scarce asset in itself.

Looking ahead, the Human-1 framework is expected to expand further to other major Indian languages such as Tamil, Bengali, and Marathi. With India's rapidly growing AI market and government policy support for indigenous language technologies, the prospects for full-duplex conversational systems in Indian languages are vast.

For the global conversational AI research community, this work also raises an important direction worth pursuing: how to make the most advanced conversational technologies truly serve multilingual, multicultural user groups, rather than remaining confined to the English-speaking world.