Blog
Read the latest news from the Chatmu blog.
An AI that listens to your mix is one that actually hears the audio coming out of your DAW, measures it, and explains what it found in numbers you can check. That is what Chatmu Desktop does in Ableton Live: hit play, and Chatmu Ears reports loudness, spectrum, dynamics, stereo image and defects, each with a unit and a target. It proposes a fix, asks before touching your set, and listens again to prove the change worked.
I'm Chema, Chatmu's founder. Before building software I spent years as a producer and sound engineer, and I still run a label with around 500 tracks. So I built this with one rule I learned in the studio: the AI measures, the engineer decides.
Chatmu is an AI agent for the music industry with its own catalog of music-industry MCP connectors. Chatmu Desktop, its app for Mac (Apple silicon and Intel) and Windows, launches on October 31, 2026, starting with Ableton Live 11 and 12; more DAWs are coming.
Updated October 2026.
Because they have no ears. Most AI assistants for DAWs work only with text. Most free Ableton MCP servers, such as ahujasid/ableton-mcp, let a chatbot control Live by text, which is genuinely useful, but they don't hear the audio. A few, like LivePilot, add an optional Max for Live analyzer that sends spectral data. Anthropic's Ableton connector for Claude answers questions from the Ableton manual. Ask a text-only assistant to fix your low end and it can move a fader to −6 dB, but it cannot tell you whether it should be −6 dB.
Mixing is a listening problem. Every decision depends on what the sound is doing right now: which instrument is hiding another, how loud the chorus is compared with the verse, whether the stereo image collapses on a phone speaker. An assistant that can only read your track names is guessing.
Chatmu Desktop puts a small plugin, Chatmu Coproducer (VST3/AU), on Ableton's Master. When you hit play, it streams what Ableton is playing to the app, and Chatmu Ears analyzes the passage. The analysis is deterministic signal processing, not a language model guessing: the same audio gives the same numbers every time. Then the agent reads those numbers, explains them in plain language, and shows a spectrogram of what it heard right in the chat, so you can see the resonance or the sibilance it is talking about.
Reading your set and listening is free of side effects. Changing anything requires you to turn on "Allow control", and destructive actions ask for approval every time.
Here is what it measures, explained the way I'd explain it to a producer in the studio.
Integrated loudness in LUFS (how loud the whole passage feels), short-term LUFS (how it moves section by section) and true peak in dBTP (the real peak after digital-to-analog conversion, which can be higher than the sample peak). Streaming services normalize playback: Spotify and YouTube play around −14 LUFS and Apple Music around −16 LUFS, and keeping true peak below about −1 dBTP gives the encoders room. Ears tells you where you are and how many dB a platform will turn you down.
Energy per frequency band, compared with a reference for your genre. Instead of "it sounds dark", you get "the 2–5 kHz band is 3 dB under the reference".
Crest factor in dB (how much your transients stick out above the average) and loudness range in LU. This is how it notices that a compressor is flattening your drums.
Correlation from −1 to +1, width as a percentage, and how many dB you lose when the mix is summed to mono, which is what happens on many phones, clubs and Bluetooth speakers.
Harsh "s" sounds, a build-up in the low-mids, and masking: when one source hides another in the same frequency range. Masking is measured per direction, because the bass hiding the kick is not the same problem as the kick hiding the bass.
Before and after a change, Ears subtracts one capture from the other. What's left is exactly what changed. If the residual is 30 dB down and only between 50 and 120 Hz, you know the edit was subtle and did what it was supposed to.
Louder always sounds better for the first few seconds. That trick fools engineers every day. So every spectral or balance comparison is made after both versions are normalized to −18 LUFS integrated. If version B wins, it wins on balance, not on volume.
| Measurement | Unit | What it tells you | Typical reference |
|---|---|---|---|
| Integrated loudness | LUFS | How loud the passage feels overall | Spotify/YouTube around −14, Apple Music around −16 |
| True peak | dBTP | Real peaks after conversion and encoding | Below about −1 dBTP |
| Band energy | dB | Tonal balance by frequency band | Genre reference, at −18 LUFS |
| Crest factor / loudness range | dB / LU | Punch and movement | Depends on genre |
| Stereo correlation / width | −1 to +1 / % | Width and mono safety | Positive correlation, small mono loss |
| Masking | 0 to 1 | Which source hides which, and where | As low as the arrangement allows |
| Null residual | dB | Exactly what an edit changed | Matches the intended change |
Because "sounds muddy" isn't actionable and "the 200–400 Hz band is +2.4 dB over the reference" is. Every value Ears returns comes with its unit and the target it is compared against, and the difference is always reported the same way: value minus target, positive means excess. You can disagree with the target. You can't misread the measurement.
The numbers below are an illustrative example, not a benchmark.
That loop of listen, propose, ask, apply and verify is the whole point. The agent never claims a fix worked because it applied it; it proves it with a second measurement.
I use plugins like these and respect them. They solve a different problem.
| Tool | What it is | Where it works | How you interact |
|---|---|---|---|
| iZotope Neutron (Mix Assistant) | Mixing plugin with an assistant | Inside the plugin on the tracks where it's inserted | Analyzes and suggests a starting point in its own processors |
| Sonible smart:EQ | Intelligent EQ plugin | Inside the plugin on each track or group | Learns the signal and applies a corrective EQ curve |
| RoEx | Online mixing service | Offline, on stems you upload | Returns a processed mix |
| Chatmu Desktop | AI agent with ears inside your DAW | Your whole Ableton set, listening from the Master | Chat: explains in units and targets, asks, applies, re-measures |
Single-plugin assistants are processors: they act inside their own plugin. RoEx works on files after the fact. Chatmu Desktop works in your live set, can act on faders, sends, sidechain and any device parameter, and tells you why in plain language. You can keep using Neutron or smart:EQ in the same session; Chatmu can even adjust their parameters.
When the mix is where you want it, Chatmu Desktop can export the song, a range or stems, and the same agent takes it from there: distribution, video and social. It hears your mix and releases it. If you live in Claude or ChatGPT instead, the Chatmu MCP brings the business side there, and the complete guide to MCP for music explains how it all connects. There's a free plan; see pricing.
Yes. With Chatmu Desktop, the Coproducer plugin on the Master sends what Ableton is playing to the app, and Chatmu Ears measures loudness, spectrum, dynamics, stereo and defects when you hit play.
Spotify normalizes playback to around −14 LUFS, Apple Music to around −16. Keep true peak below about −1 dBTP. Chatmu reports your integrated LUFS and how much each platform would turn you down.
No. Listening and reading are free; changes need "Allow control", and destructive actions ask for approval each time.
No. It's AI inside your DAW. You decide, everything stays editable and the stems are yours.
Chatmu Desktop starts with Ableton Live 11 and 12 on Mac and Windows; more DAWs are coming.
AI for the industry. Humans for the music.
