Moshi — Real-Time AI Voice Conversation Engine

What it is

Moshi is an open-source real-time voice AI system built by Kyutai. It supports full-duplex speech conversation, meaning both parties can speak and listen simultaneously, with approximately 200ms latency. It includes emotion recognition and runs on-device without requiring cloud APIs.

Moshi targets developers building voice interfaces, conversational assistants, and accessibility tools who need low-latency, privacy-preserving voice interaction.

How it saves time or tokens

Moshi processes speech on-device, eliminating round-trip latency to cloud speech APIs. The full-duplex architecture removes turn-taking overhead typical of voice assistants. Estimated token usage for this workflow is around 3,900 tokens.

How to use

1. Install and start the server:

pip install moshi
python -m moshi.server

1. Open http://localhost:8998 in your browser.
2. Start speaking. Moshi responds in real time with full-duplex audio.

Example

# Install Moshi
pip install moshi

# Start the voice server
python -m moshi.server

# Open browser to http://localhost:8998
# Speak naturally — Moshi responds with ~200ms latency

The browser interface handles audio capture and playback. No additional microphone setup is needed beyond browser permissions.

Related on TokRepo

• AI Tools for Voice — More voice AI tools and engines
• Local LLM Providers — Run AI models locally alongside Moshi

Key considerations

When evaluating Moshi for your workflow, consider the following factors. First, assess whether your team has the technical prerequisites to adopt this tool effectively. Second, evaluate the maintenance burden against the productivity gains. Third, check community activity and documentation quality to ensure long-term viability. Integration with your existing toolchain matters more than feature count alone. Start with a small pilot project before rolling out across the organization. Monitor resource usage during the initial adoption phase to identify bottlenecks early. Document your configuration decisions so team members can onboard independently.

Common pitfalls

• On-device processing requires adequate GPU or CPU resources; low-end machines may experience higher latency.
• Browser audio permissions must be granted for the microphone to work.
• Emotion recognition accuracy varies by language and accent; primary testing has been on English speech.