Audio AI GUIDE
Mimi Streaming Audio Codec
Mimi is Kyutai’s neural audio codec used with Moshi.
Listen to this guideAI-narrated · ~2 min
On this page2 min read
Overview
It converts 24 kHz audio into discrete representations at 12.5 frames per second, enabling speech models to predict and decode audio tokens. The paper describes a semantic codebook distilled from WavLM and acoustic codebooks for reconstruction; the specific rates and latency belong to the published configuration.
Deep Dive
The Moshi paper introduces Mimi as a streaming neural codec operating on 24 kHz audio with a 12.5 Hz token frame rate and 1.1 kbps bandwidth in the reported setup. The encoder produces a semantic codebook distilled from WavLM and acoustic codebooks that support reconstruction, using split residual vector quantization. Causal masking in the encoder/decoder Transformers supports streaming inference. The paper reports an 80 ms frame size; it should not be read as a guaranteed end-to-end conversational response time. Mimi supplies audio tokens for Moshi’s audio-text modeling, but codec tokens are not ordinary text tokens.
Real-World Implementation
Use a codec to represent incoming speech as discrete tokens that a speech-text model can process.
Decode predicted audio tokens back into a waveform for speech output.
Compare token rate and bitrate when assessing autoregressive sequence length and codec fidelity.
Inspect the split semantic/acoustic codebooks described in the Moshi report.
The Future of Mimi Streaming Audio Codec
Neural codecs provide one interface for speech models to encode and generate audio. Future work may adjust token rates, codebook allocation, and streaming latency; each trade-off must be evaluated for the intended audio quality, hardware, and dialogue task.
Putting it into practice
Why it matters
- 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.
Watch out for
- 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.
Next steps
- 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
Frequently asked questions
What is Mimi Streaming Audio Codec?
Mimi is Kyutai’s neural audio codec used with Moshi. It converts 24 kHz audio into discrete representations at 12.5 frames per second, enabling speech models to predict and decode audio tokens. The paper describes a semantic codebook distilled from WavLM and acoustic codebooks for reconstruction; the specific rates and latency belong to the published configuration.
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 information does Mimi’s first semantic codebook aim to represent?
The Moshi paper describes distilling WavLM representations into Mimi’s semantic codebook; it does not claim that this codebook is a literal transcript or exact speaker-identity record.
Why is Mimi described as 'streaming' or 'causal'?
Mimi’s causal encoder/decoder can process frames as audio arrives rather than requiring the full utterance. The paper’s 80 ms figure is a codec frame interval, not a guaranteed end-to-end dialogue latency.
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.
Keep learning
Related guides
More guides picked for this topic