DEV Community

Ai Game Changer Pro
Ai Game Changer Pro

Posted on Originally published at aigamechangerpro.blogspot.com

Haru-4B: Talk to a Free AI That Listens and Speaks at the Same Time — Full Setup Guide

Every voice AI you have used works the same way: you talk, it waits, it thinks, then it answers. Real conversation does not work like that. Real conversation is full of interruptions, half-finished sentences and people talking over each other — and now there is a free, open-source model that can do the same.

What Haru-4B Is

Haru-4B is a full-duplex speech-to-speech model from higurashi-ai. "Full-duplex" means it listens and speaks at the same time, just like a phone call. You can interrupt it mid-sentence, talk over it, and it responds in about 65 milliseconds — around the same time a human needs to react. It is released under the permissive Apache 2.0 licence.

The model has 4 billion parameters and supports 9 languages: English, Japanese, Chinese, Korean, French, German, Spanish, Portuguese and Italian. It can sense laughter and changes in tone, and it can be given a character and a voice of its own.

What You Need to Run It

  • Your own GPU: an Nvidia card with at least 12 GB of VRAM. The author recommends an RTX 3090, 4090 or 5090, or a Mac with Apple Silicon.
  • Free, no install: a ready-made Google Colab notebook on the project's GitHub page.

How to Try It for Free on Colab

  1. Open the GitHub repository: github.com/higurashi-ai/Haru
  2. Open the Colab notebook listed there.
  3. Run the cells from the top. The first run downloads the model and takes a few minutes.
  4. Click the Gradio link and allow microphone access.
  5. Speak naturally. You can interrupt it, change the subject mid-sentence, or speak over it.

How to Give It a Character and a Voice

You can change the model at the level of the AI itself. Start with the default voice, then give it a role (a teacher, a travel guide, a customer-service agent) and a personality (helpful, witty, patient). The system prompt is the place to do it — describe the character, the language to use, the level of formality, and the voice it should answer in.

The project also has tools for cloning a voice from a short sample and creating custom voice profiles, so the same model can sound like a different person for each role.

The Limits

  • It is not built for coding or long-form writing.
  • It has no long-term memory beyond the conversation.
  • Its knowledge may be slightly behind a flagship model.
  • It can still repeat itself, ramble or speak oddly on long or complex topics.
  • On free Colab, sessions end and you lose the conversation, and the model sometimes drops audio.

How to Use It From Python

A Python wrapper is included in the repository. With vllm and the haru package installed, you can use the model in the background, in your own scripts and in your own apps, with full control over input, output, voice and character.

The model is also available on Hugging Face. If you prefer a ready-made app, use the Colab notebook instead.

How Haru-4B Compares

The model runs in about 65 milliseconds — compared with about 232 milliseconds for GPT-4o Realtime API, and about 320 milliseconds for Moshi. The open-source model Sesame CSM has no latency figure listed.

What It Can Be Used For

Voice assistants and customer service, interactive games and virtual characters, education and language learning, accessibility tools, and hobby projects.

The Bottom Line

For the first time, a free, open-source AI can hold a real voice conversation: talk, interrupt, talk over it, and get an answer in about 65 milliseconds, in 9 languages, on a 12 GB GPU or free on Colab. The future of voice AI may not be built behind a subscription — it may be built in the open.

Originally published on Ai Game Changer Pro.

Top comments (0)