Is Fun-Audio-Chat free?

Yes, it is an open-source project available on GitHub and Hugging Face, allowing developers to use and modify it freely.

What is the difference between Fun-Audio-Chat and FunASR?

FunASR is primarily for recognizing speech (turning audio into text), while Fun-Audio-Chat is a full interaction model that understands the audio, thinks, and generates a spoken response (handling the entire conversation).

Can it detect emotions?

Yes, "Voice Empathy" is a core feature, allowing the model to understand if a user is happy, sad, or angry and adjust its response accordingly.

It was built by the FunAudioLLM Team, which is part of the research division at Alibaba (similar to the team behind Qwen and MinMo).

Fun-Audio-Chat

A Large Audio Language Model built for natural, low-latency voice interactions.

From Alibaba Group

Founder: FunAudioLLM Team (Alibaba Group/Tongyi Laboratory)

Audio & Music

Dec 2025

20K+

Free

Pricing Model

Free

Starting Price

$0/Month

About This AI

Fun-Audio-Chat is a state-of-the-art Large Audio Language Model (LALM) designed to enable highly natural, low-latency voice conversations.

Developed by the same team behind FunASR, it introduces “Dual-Resolution Speech Representations” to process audio efficiently, cutting GPU usage by nearly 50% without sacrificing quality.

It excels at understanding spoken questions, following complex voice instructions, and even detecting and responding with “voice empathy,” making it one of the most capable open-source engines for building human-like voice agents.

Audio Editor Chatbot

Pricing

Pricing Model

Free

Starting Price

$0/Month

Key Features

Dual-Resolution Architecture: Uses a split processing method (5Hz backbone + 25Hz head) to reduce computational load while keeping high audio fidelity.
Voice Empathy: capable of detecting the user's emotional tone and responding with appropriate emotional inflection.
Speech Function Calling: Can trigger external tools or functions directly via voice commands (e.g., "Set an alarm").
Speech Instruction Following: accurately executes complex multi-step instructions given verbally.
Low Latency: Optimized for near-instantaneous response times, essential for fluid conversation.
Core-Cocktail Training: A specialized training method that preserves the text capabilities of the underlying LLM while adding audio skills.

Pros

Significant efficiency gains (lighter on GPU memory).
Strong performance in "VoiceBench" and "OpenAudioBench" evaluations.
Open-source and customizable.
Bridges the gap between text intelligence and audio nuance.

Cons

Requires technical setup (Python, PyTorch) to run.
Hardware intensive for training (requires 4x 80GB GPUs for training, though inference is lighter).
Primarily a backend/middleware tool, not a consumer app with a UI.

FAQs

Is Fun-Audio-Chat free?
Yes, it is an open-source project available on GitHub and Hugging Face, allowing developers to use and modify it freely.
What is the difference between Fun-Audio-Chat and FunASR?
FunASR is primarily for recognizing speech (turning audio into text), while Fun-Audio-Chat is a full interaction model that understands the audio, thinks, and generates a spoken response (handling the entire conversation).
Can it detect emotions?
Yes, “Voice Empathy” is a core feature, allowing the model to understand if a user is happy, sad, or angry and adjust its response accordingly.
Who built this tool?
It was built by the FunAudioLLM Team, which is part of the research division at Alibaba (similar to the team behind Qwen and MinMo).