Human-1: a full-duplex conversational system in Hindi

A cascaded voice pipeline is a sequence of irreversible commitments: the endpointer decides the turn has ended, the recogniser commits to a transcript, the language model commits to a reply, and only then does synthesis begin. Overlap, interruption and backchannelling — the parts of conversation that make it feel like conversation — have nowhere to live in that sequence.

Full-duplex architectures remove the sequence, but the published work assumes English: English tokenizers, English turn-taking behaviour, and large English conversational corpora. For Hindi none of the three existed, and spontaneous Hindi is heavily code-mixed, so a tokenizer built for one script handles the inner-monologue text stream badly.

We adapted the Moshi/Mimi architecture directly rather than assembling a cascade, so that the user stream and the model stream are both always on and overlap is represented in the model rather than resolved before it.

Two pieces of that adaptation were specific to Hindi. A 32k-token SentencePiece tokenizer was trained for the text stream, sized to cover Devanagari and Latin code-mixing without exploding sequence length. And the training data was real-world stereo conversation — 26,000 hours of it — rather than synthesised or read speech, because turn-taking, hesitation and backchannel behaviour cannot be recovered from prompted recordings.

Training ran as a distributed multi-GPU pipeline across an 8× NVIDIA H100 cluster, which set the practical limits on corpus preprocessing throughput as much as on model size.

In perceptual quality and turn-taking baseline tests, Human-1 reached a human naturalness score of 4.10 out of 5.00, and 66.9% of its generations tied directly with genuine human speech.

The system and the framework around it are released open-source, with the preprint documenting the architecture adaptation, tokenizer, and evaluation protocol.