README.md

---
license: cc-by-4.0
language:
  - hi
tags:
  - moshi
  - speech-to-speech
  - hindi
  - conversational-ai
  - audio
  - full-duplex
  - duplex-dialogue
  - indian-languages
base_model: kyutai/moshiko-pytorch-bf16
pipeline_tag: audio-to-audio
---

# Human-1: A Full-Duplex Conversational Model for Hindi
**🎙️ [Try the live demo →](https://ai.joshtalks.com/research/human-1)** | **📄 [Paper →](https://arxiv.org/pdf/2604.23295v1)**

Human-1 by Josh Talks is the first full-duplex spoken dialogue model for Hindi, built by adapting [Kyutai's Moshi](https://github.com/kyutai-labs/moshi) architecture. It enables real-time, natural Hindi conversation with support for interruptions, overlaps, backchannels, and natural turn-taking — trained on 26,000 hours of real spontaneous Hindi conversations from 14,695 speakers.

<p align="center">
  <img src="hindi_moshi_architecture.svg" alt="Hindi-Moshi Architecture" width="480"/>
</p>

## Model Details

| | |
|---|---|
| **Developed by** | Bhaskar Singh, Shobhit Banga, Pranav Sharma — [JoshTalks](https://joshtalks.com) |
| **Base model** | [kyutai/moshiko-pytorch-bf16](https://huggingface.co/kyutai/moshiko-pytorch-bf16) |
| **Language** | Hindi (hi) |
| **Model type** | Full-duplex speech-to-speech dialogue |
| **Format** | SafeTensors (fp32) |
| **Tokenizer** | Custom Hindi SentencePiece (32,000 vocabulary) |
| **Audio codec** | Mimi (frozen, 12.5 Hz, 1.1 kbps) |
| **License** | CC-BY-4.0 |

## What was changed from base Moshi

The original English SentencePiece tokenizer was replaced with a Hindi SentencePiece model (32,000 vocabulary) trained on a large Hindi text corpus. This required reinitialisation of three vocabulary-dependent parameter groups:

- `text_emb` — text token embedding in the Temporal Transformer
- `depformer.emb.0` — text token embedding in the Depth Transformer
- `text_linear` — text output projection layer

All audio processing components (Mimi codec) and remaining transformer weights retain their pre-trained values. Mimi generalises to Hindi without retraining (STOI: 0.878, PESQ: 2.55).

For full architecture details, see the [Moshi paper](https://arxiv.org/abs/2410.00037).

## Training

### Data

The model was trained on a purpose-built corpus of **26,000 hours** of real Hindi spontaneous conversations — to our knowledge, the largest conversational speech corpus for any Indian language.

| Characteristic | Value |
|---|---|
| Total duration | 26,000 hours |
| Unique speakers | 14,695 |
| Recording type | Spontaneous, unscripted conversations |
| Channels | Stereo (separate per speaker) |
| Quality control | Trained annotators + manual checks |

The stereo recording format with separate speaker channels enables direct learning of turn-taking, overlaps, and backchannels from natural interactions — without requiring artificial speaker diarisation.

### Two-stage training recipe

**Stage 1 — Pre-training** on the full 26,000-hour corpus. Learning rate of 3×10⁻⁵ (matching original Moshi pre-training). AdamW with β₁=0.9, β₂=0.95, weight decay 0.1. Effective batch size of 64 (\~2.9 hours of audio per update). Trained for 1 epoch (\~10,000 steps) in approximately 13 hours on 8× NVIDIA H100 80GB GPUs.

**Stage 2 — Fine-tuning** on ~990 hours of curated high-quality conversational data. Split learning rates: 2×10⁻⁶ for the Temporal Transformer, 4×10⁻⁶ for the Depth Transformer. Optimal checkpoint selected at step 4,812 based on minimum total validation loss (3.370).

### Training infrastructure

8× NVIDIA H100 80GB GPUs with bf16 mixed precision.

## Evaluation

### Perplexity

Measured using Sarvam-1 (2B) on Whisper-v3 transcriptions of generated speech.

| Temperature | PPL ↓ |
|---|---|
| Ground-truth | 237.1 |
| Human-1 (τ=0.8) | 356.9 |
| Human-1 (τ=0.9) | 467.1 |
| Human-1 (τ=1.0) | 640.6 |

### Human Evaluation

130 evaluators completed 2,125 rating tasks comparing human speech with model responses. Each instance contained two audio samples (Voice A: Human, Voice B: Model) rated on 5-point Likert scales for naturalness and clarity.

**Perceptual quality:**

| Metric | Human Score | Model Score | Human Preferred | Model Preferred | Tie |
|---|---|---|---|---|---|
| Naturalness | 4.55 | 4.10 | 30.0% | 3.1% | 66.9% |
| Clarity | 4.05 | 3.04 | — | — | — |

Generated speech achieves high perceptual quality, with naturalness scores approaching human speech and most pairwise comparisons resulting in ties.

**Conversational rubric evaluation:**

Evaluators also assessed conversational quality using three binary rubric questions measuring whether generated responses behave like natural conversational speech.

| Rubric | Pass Rate |
|---|---|
| Human-like interaction | ≈85% |
| Appropriateness (response follows prompt) | ≈53% |
| Completion (response forms a complete reply) | ≈42% |

While the model frequently produces speech that sounds human-like, maintaining contextual relevance and producing fully complete conversational responses remains an ongoing challenge.

### Turn-Taking Analysis

Temperature τ=0.9 produces turn-taking dynamics closest to ground-truth.

| Model | τ | IPU/min | Pause | Gap | Overlap |
|---|---|---|---|---|---|
| Ground-truth | — | 35.30 | 10.49 | 8.51 | 3.03 |
| Human-1 | 0.8 | 23.12 | 9.16 | 6.77 | 1.67 |
| Human-1 | 0.9 | 29.14 | 9.24 | 8.54 | 4.30 |
| Human-1 | 1.0 | 38.90 | 11.67 | 8.10 | 9.68 |

## Conversation Style

Human-1 is trained on **topic-driven conversations** - real dialogues where two speakers discuss a subject naturally, with backchannels, interruptions, and organic turn-taking.

After an initial introduction, the model will typically **propose a topic and steer the conversation toward it**, preferring structured discussion over open-ended chitchat. Users can also **introduce their own topic** - the model will pick it up and engage in a focused discussion around it. This is an intentional design choice - the training data consists of real conversations where speakers engage in focused, in-depth discussions on assigned topics.

This makes the model particularly well-suited for **domain-specific conversational applications**. Our key finding is that the model's ability to stay on-topic emerges naturally from the structure of the training data alone - without any explicit prompting, reward shaping, or guardrails. This suggests that with sufficient hours of domain-specific conversational data, this approach can produce models that learn the conversational norms of virtually any domain - customer support, healthcare consultations, language tutoring, sales, therapy, and more - opening a direct path from curated conversations to deployable, real-world voice agents. Exploring this is an active direction of our future work.

## Files

```
├── model.safetensors                              # Human-1 LM weights
├── tokenizer-e351c8d8-checkpoint125.safetensors   # Mimi audio codec (frozen, from Moshi)
├── tokenizer_hindi.model                          # Hindi SentencePiece tokenizer
├── tokenizer_hindi.vocab                          # Vocabulary reference
├── hindi_moshi_architecture.svg                   # Architecture diagram
└── README.md
```

## Quick Start

### 1. Install uv

```bash
curl -LsSf https://astral.sh/uv/install.sh | sh
source $HOME/.local/bin/env
```

### 2. Create project and install dependencies

```bash
uv init human-1 && cd human-1
uv python install 3.12
uv python pin 3.12
uv add moshi huggingface_hub
```

### 3. Download the model

```bash
uv run huggingface-cli download JoshTalksAI/Human-1 --local-dir ./weights
```

### 4. Run the server

```bash
uv run -m moshi.server \
    --moshi-weight ./weights/model.safetensors \
    --mimi-weight ./weights/tokenizer-e351c8d8-checkpoint125.safetensors \
    --tokenizer ./weights/tokenizer_hindi.model
```

## Intended Use

The model is intended for research in full-duplex spoken dialogue systems for Hindi and Indian languages. It can be used as a conversational agent for casual Hindi conversations.

## Limitations

- Trained primarily on Hindi conversational speech. Performance on other languages or domains is not guaranteed.
- Inherits limitations from the base Moshi architecture regarding audio quality at 1.1 kbps bitrate.
- Hindi text tokens are sparser relative to audio (~75% PAD ratio vs. 65% in English) due to Devanagari encoding more phonemic content per token.
- Not intended for impersonation or any malicious use.
- This model is for research purposes. We do not recommend it for providing advice or performing any professional duty.

## Citation

```bibtex
@article{singh2026human1,
  title   = {Human-1 by Josh Talks : A Full-Duplex Conversational Modeling Framework in Hindi using Real-World Conversations},
  author  = {Bhaskar Singh and Shobhit Banga and Pranav Sharma},
  year    = {2026},
  institution = {JoshTalks}
}
```

## Acknowledgments

Built on [Moshi](https://github.com/kyutai-labs/moshi) by [Kyutai](https://kyutai.org/). We thank the 14,695 speakers who contributed to the Hindi conversational corpus.