nu-dialogue/j-moshi

J-Moshi: A Japanese Full-duplex Spoken Dialogue System

What it solves

J-Moshi is a full-duplex spoken dialogue system designed for the Japanese language. It addresses the challenge of creating natural, real-time voice interactions that mimic human-to-human conversation, specifically enabling features like overlapping speech and backchanneling (aizuchi) in real-time.

How it works

The system is based on the 7B parameter Moshi model from Kyutai Labs. It was developed through additional training on large-scale Japanese spoken dialogue data, including corpora such as J-CHAT, Japanese Callhome, and internal chat and consultation dialogue corpora. A specialized version, J-Moshi-ext, further incorporates synthetic data generated via Multi-stream TTS to enhance performance.

Who it’s for

Researchers and developers working on Japanese voice AI, spoken dialogue systems, and natural human-computer interaction (HCI).

Highlights

  • Full-duplex communication: Enables real-time, bidirectional voice interaction with natural turn-taking.
  • Curation of Japanese data: Trained on a diverse set of spoken and text-based dialogue corpora.
  • Two model variants: Offers a standard version and an extended version (J-Moshi-ext) using synthetic data.
  • HuggingFace integration: Models are available for easy deployment via the PyTorch implementation of Moshi.

Related

  • Project
    flyteorg/flytekit

    Flytekit Python is the official SDK for Flyte, letting you write, test, and deploy AI/ML workflows in pure Python using `@task` and `@workflow` decorators. Install with `pip install flytekit`, follow the quick‑start guide, and extend via plugins or contribute through the provided docs.

  • Project
    aa0101181514/tw-legal-rag

    tw‑legal‑rag is an open‑source Python CLI that connects to a hosted semantic‑search service containing 22 million+ Taiwanese judgments, statutes and administrative rulings. It lets you search with natural language, fetch exact case numbers, bundle results (including excerpts, case‑history and citation whitelists) into a JSON file, and run a deterministic citation‑check against LLM‑generated answers. The tool itself never calls an LLM and works without API keys, serving as a retrieval‑augmented generation (RAG) front‑end for legal AI applications.

  • Project
    OpenVoiceOS/ovos-installer

    A streamlined installer for Open Voice OS, enabling users to deploy a privacy-focused, open-source voice assistant on Raspberry Pi, Linux, or Mac hardware.

  • Project
    csyangwen/dsh-memory-evolve

    dsh‑memory‑evolve is a DSH plugin that gives the AI persistent cross‑session memory, todo/project logs, Git‑based memory sync, multi‑agent orchestration (internal sub‑sessions and external AI providers), a searchable infinite canvas for files, optional reviewer sessions, and mobile‑friendly UI with notifications. It lets the AI act as a long‑term partner and team coordinator while keeping the user in control of all writes.