The model can listen and speak simultaneously, enabling AI conversations that mirror how people actually talk
Trained on 26,000 hours of real Hindi conversations across 14,695 speakers, capturing interruptions, pauses and natural dialogue flow
Demonstrates how large-scale conversational datasets can unlock next-generation voice AI for Indian languages
New Delhi, March 11, 2026: Josh Talks, a company building large-scale speech datasets and evaluation infrastructure for AI,, today announced the development of the world’s first full-duplex conversational AI model in Hindi, an AI system capable of listening and speaking at the same time, enabling voice interactions that more closely resemble natural human conversation.
Unlike traditional voice assistants that operate in a turn-based format, where users must finish speaking before the system responds, Josh Talks model can listen and speak simultaneously. This allows it to acknowledge users mid-sentence, respond in real time, and handle interruptions without breaking the flow of conversation.
Shobhit Banga, Co-Founder, Josh Talks, said, “Voice AI has learned to recognise speech and generate speech. What it has not yet learned well is how to participate in conversation. This research shows that when models are trained on large-scale natural dialogue, they can begin to learn the rhythm of how people actually speak to each other. By building this infrastructure using Hindi conversations, we have demonstrated that full-duplex conversational systems can be developed beyond English. The mission of Josh Talks AI is to make machines talk like humans. The 26,000-hour conversational dataset used for this work represents the first milestone in a broader effort to build a one-million-hour corpus of natural conversations, one of the largest datasets ever created for conversational voice AI. ”
Most voice assistants today struggle with the natural dynamics of human conversation. Real dialogue rarely follows a rigid pattern of one person speaking while the other waits. People interrupt, pause, acknowledge each other mid-sentence and use brief cues such as “haan,” “achha,” or “hmm” to signal understanding.
Traditional AI systems often fail to capture this rhythm because they process speech sequentially. Users speak, the system processes the input, and only then does it respond. This delay makes interactions feel mechanical and limits how naturally people can communicate with AI.
Human-1 by Josh Talks addresses this challenge by enabling an AI system that can process incoming speech while generating its own responses, closer to how humans participate in conversations. The result is a more fluid interaction where the system can adapt dynamically to the flow of dialogue.
To build the model, the Josh Talks team trained it on 26,000 hours of natural two-person Hindi conversations involving 14,695 unique speakers. Unlike traditional speech datasets that rely on scripted or structured inputs, the dataset captures how people actually communicate in everyday situations, including overlapping speech, interruptions, pauses, conversational acknowledgements and spontaneous reactions.
This type of data is critical for advancing voice AI. While earlier AI models were trained primarily on text or clean speech samples, real-world conversations are far more complex and unpredictable. Capturing these patterns requires large-scale datasets that reflect authentic human dialogue rather than carefully controlled recordings. By learning from these real interactions, the system begins to understand not just words, but also the timing, rhythm and responsiveness that make conversations feel natural.
To encourage further research and experimentation in conversational voice AI for Indian languages, Josh Talks has open sourced the model developed through this work.
Josh Talks has open sourced the Hindi duplex conversational model to encourage further research on conversational voice AI for Indian languages. By making the model publicly available, the team hopes to enable researchers and developers to experiment with full-duplex dialogue systems and build upon the approach. The release also highlights the role of large-scale conversational datasets in enabling more natural and responsive voice AI systems. While the underlying conversational dataset remains proprietary, the model release allows the research community to explore full-duplex conversational architectures for Hindi.
The development places Josh Talks among a small group of organisations globally exploring full-duplex conversational AI architectures. What makes this breakthrough particularly significant is that it brings this capability to Hindi, a language spoken by hundreds of millions of people but historically underrepresented in advanced AI systems. As voice increasingly becomes the primary interface for technology, especially in markets like India where typing is not always the preferred mode of digital interaction, enabling natural conversation in local languages is critical for expanding digital access.
While the current model focuses on Hindi, the underlying methodology has been designed to be replicable across languages. India’s linguistic diversity presents both a challenge and an opportunity for AI development. Building systems that understand how people speak across different languages requires large volumes of conversational data and deep contextual understanding of how dialogue unfolds in real life.The Josh Talks Duplex Model demonstrates that by collecting and training on large-scale conversational datasets, similar systems could eventually support natural voice interaction across multiple Indian languages.
The research highlights a growing shift in the AI ecosystem, from training models on text and isolated speech samples to building conversation-scale datasets that capture the complexity of human interaction. These datasets help AI systems learn critical elements of communication such as timing, tone, responsiveness and conversational context, capabilities that will define the next generation of voice technology.
By focusing on real conversational behaviour, Human-1 by Josh Talks moves beyond the limitations of traditional voice systems and points to a future where AI adapts to the way humans communicate, rather than forcing humans to adapt to machines. With the introduction of the world’s first full-duplex conversational AI model in Hindi, Josh Talks marks an important step toward making voice technology more natural, inclusive and accessible for millions of speakers of Indian languages.
Josh Talks AI is a sister company of Josh Talks Media, one of India’s largest vernacular storytelling platforms.
Josh Talks AI focuses on building large-scale conversational speech datasets, benchmarks, and evaluation infrastructure to support the development of voice AI for Indian languages. The company works with leading AI labs and research teams to help train, test, and improve speech and conversational AI systems for real-world usage in India.
To explore the model and learn more, visit: https://ai.joshtalks.com/research/human-1