Mimi Streaming Audio Codec
Mimi is a neural audio codec that compresses speech into a tiny stream of discrete tokens in real time, so AI models can listen and speak with very low latency.
Overview
It is the audio backbone behind Kyutai's Moshi voice model.
Deep Dive
Mimi, released by the French lab Kyutai in 2024, is a neural codec that turns 24 kHz audio into a stream of discrete tokens at roughly 1.1 kbps and only 12.5 tokens per second. It uses an encoder-decoder with residual vector quantization (RVQ), splitting tokens into a 'semantic' first level distilled from a self-supervised speech model (WavLM) plus several 'acoustic' levels that capture voice texture. Crucially it is fully streaming and causal: it emits tokens as audio arrives rather than waiting for a full clip, with about 80 ms of latency. This lets a language model treat speech like text tokens, enabling Moshi to converse in full duplex while keeping reconstructed audio intelligible and natural.
Technical Insight
Mimi's trick is a split-RVQ scheme. The first codebook is trained with a distillation loss to match embeddings from WavLM, forcing it to carry phonetic 'meaning,' while parallel acoustic codebooks reconstruct waveform detail. A Transformer operates inside the bottleneck, and an adversarial (GAN) loss on the decoder sharpens output quality. Causal convolutions keep everything streaming, so latency stays near 80 ms.
Strategic Impact
Access and reach
It improves accessibility through transcription, narration, and voice interfaces.
Cost and budget
Media teams can ship polished audio faster with smaller budgets.
Speed and scale
Customer-facing systems can process spoken interactions at larger scale.
The Future of Mimi Streaming Audio Codec
Expect codecs like Mimi to become the standard interface between audio and large language models, pushing real-time voice assistants toward sub-100 ms response times. Research is driving token rates even lower while preserving speaker identity, emotion, and music. Because Kyutai open-sourced Mimi and Moshi, it is likely to seed many open speech-to-speech systems, on-device assistants, and ultra-low-bandwidth voice communication tools.
Real-World Implementation
Powering Kyutai's Moshi full-duplex voice assistant so it can listen and talk simultaneously
Streaming speech tokens into a language model for real-time speech-to-speech translation
Ultra-low-bitrate voice calls (~1.1 kbps) for poor or congested network conditions
Tokenizing audio for generative speech and text-to-speech pipelines that reason over sound like text
Risks & Guardrails
Voice misuse and impersonation risks increase when consent is missing.
Accuracy can drop across accents, dialects, or noisy environments.
Synthetic audio can be mistaken for authentic speech without clear labeling.
Implementation Roadmap
Obtain explicit consent for voice capture, cloning, and reuse.
Test quality across diverse speakers and background conditions.
Define when a human must review or approve outputs.
Label synthetic audio and keep provenance records for accountability.
Keep Exploring
Free newsletter
Keep up with AI in 3 minutes a day
One short email each weekday with the three AI stories that actually matter. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Mimi Streaming Audio Codec quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Next guide
Neural Audio Codecs
Frequently asked questions
What is Mimi Streaming Audio Codec?
Mimi is a neural audio codec that compresses speech into a tiny stream of discrete tokens in real time, so AI models can listen and speak with very low latency. It is the audio backbone behind Kyutai's Moshi voice model.
What lab released Mimi and the Moshi voice model?
Mimi and Moshi were released in 2024 by the French research lab Kyutai, which open-sourced both.
Roughly how many tokens per second does Mimi emit for speech?
Mimi compresses 24 kHz audio down to only about 12.5 tokens per second at roughly 1.1 kbps.
What does the first ('semantic') codebook in Mimi capture?
The first RVQ level is trained via distillation from WavLM to carry semantic/phonetic content, while later levels add acoustic detail.
Why is Mimi described as 'streaming' or 'causal'?
Mimi processes audio causally and outputs tokens incrementally, giving about 80 ms latency suitable for live conversation.
Which quantization technique does Mimi use to encode audio?
Mimi uses residual vector quantization, stacking multiple codebooks so each adds finer detail to the previous one.