AI
First full-duplex Hindi dialogue system handles natural conversation patterns like interruptions and overlaps
Researchers have built the first open-source conversational AI system for Hindi that learns from real speech data to model authentic dialogue behaviors.
Researchers working with Josh Talks have unveiled Human-1, a speech dialogue system purpose-built for Hindi that can capture the messiness of real conversations—including when people talk over each other, jump in mid-sentence, or make brief acknowledgment sounds.
What sets it apart#
Unlike most conversational AI systems, which tend to process language in clean, turn-based exchanges, full-duplex systems allow for simultaneous speech. This enables the model to learn authentic dialogue patterns that humans produce naturally: interruptions, overlaps, and backchannels (like "haan" or "hmm"). Such capabilities have remained largely underdeveloped for Indian languages until now.
How it was built#
The team adapted Moshi, an advanced duplex speech architecture, by swapping in a custom Hindi tokenizer and retraining the model on an enormous corpus: 26,000 hours of spontaneous conversations sourced from 14,695 speakers, each recorded on separate channels. This channel separation was critical—it allowed the system to learn turn-taking and overlap dynamics directly from natural interactions rather than synthetic data.
To support Hindi text generation, researchers replaced the original English tokenizer and reset text-vocabulary parameters while keeping the pre-trained audio components intact. The training followed a two-stage approach: large-scale pre-training on the full conversational dataset, then refinement using 1,000 hours of curated dialogue samples.
Validation and impact#
Performance was measured using dialogue continuation tasks, where the model extends conversational exchanges, evaluated with both automatic metrics and human judges. The results show the system generates natural and coherent Hindi conversation with authentic dialogue behaviors.
The researchers position this as a foundation for real-time duplex dialogue systems across Hindi and other Indian languages—areas where such technology has been largely absent.
