Samsung Patents a Way to Pull Apart Mixed Audio Into Separate Sound Objects
Imagine recording a video at a concert and later being able to turn down the crowd noise while keeping the music crisp. Samsung's latest patent describes a system that tries to make exactly that kind of audio separation work on a device itself, without sending your audio to the cloud.
How Samsung's audio separation actually works
When you record a video in a noisy room, every sound gets baked into one flat audio track. Voices, music, background hum, and everything else are mixed together. Separating them after the fact is something AI can do, but it normally takes a lot of computing power.
Samsung's patent describes a way to make that separation process lighter and faster. Before the AI model even looks at the audio, the system converts the sound into a compressed form that strips out redundant frequency information. Think of it like resizing a huge photo before uploading it: the image still looks fine, but the file is much smaller and faster to handle.
The result is that the AI model has less data to chew through, which means the whole separation process can run on a phone or tablet without draining the battery or causing a lag. Each sound object, like a voice or a piece of music, gets its own isolated audio track.
How the frequency compression feeds the separation model
The patent describes a two-stage pre-processing pipeline that feeds a neural-network-based audio object separation model.
Stage one takes incoming audio and converts it from a raw time-domain waveform into the frequency domain using something like a Short-Time Fourier Transform (STFT), which breaks audio into a grid of frequency "bins" over time. This produces what the patent calls "first frequency domain data."
Stage two compresses that frequency grid down to a smaller number of bins using a "designated conversion algorithm." The patent doesn't lock in one specific algorithm, but this step mirrors techniques like mel-scale filterbanks (which group frequency bins the way human hearing does, discarding detail that the ear wouldn't notice anyway).
The compressed data is then passed into the object separation model, which outputs a set of "masks", essentially per-frequency filters, one for each sound source the model has identified. Applying those masks to the original audio reconstructs each isolated sound object as its own audio stream.
- Fewer frequency bins mean a smaller input tensor (the data grid the AI processes)
- A smaller input means faster inference and lower memory use on mobile hardware
- Each mask targets one sound object: a voice, instrument, background noise, etc.
What this means for Samsung's audio and video tools
For Samsung, this kind of on-device audio separation is a practical building block for features across Galaxy phones, tablets, and earbuds. Video editing tools that let you remix audio after recording, hearing-aid-style features that isolate voices, or live call-enhancement that strips background noise all depend on fast, efficient sound separation.
The key design choice here is the compression step before the AI model runs. By reducing the input size up front, Samsung can potentially run this on mid-range devices, not just flagship hardware. That matters because it would let the company ship the feature broadly rather than as an exclusive for top-tier models.
This is a solid, practical engineering patent rather than a flashy concept. The compression-before-inference approach is a well-understood efficiency trick, and Samsung is applying it to a real problem that affects every phone camera and audio recording. It won't make headlines on its own, but it's exactly the kind of work that makes AI audio features actually ship on affordable hardware.
Which company should we read for you?
We track 17 companies here. Pro is the same weekly breakdown for any company you choose, delivered privately. Type a name and we'll scope it and send you a quote.
Get one Big Tech patent every Sunday
Plain English, intelligent commentary, no hype. Free.
Editorial commentary on a publicly published patent application. Not legal advice.