Liquid State Machine for Audio Denoising
A Liquid State Machine using spiking neural networks to recognise temporal audio patterns and filter unwanted noise from sound signals.
- Type
- Spiking Neural Networks · Neuromorphic Computing
- Role
- AI Engineer / Researcher
- Contribution
- Research · Network design · Audio processing
The problem
Background noise degrades speech and audio everywhere from calls to recordings, and removing it well depends on recognising patterns that unfold over time rather than in a single instant. Conventional approaches handle this with large networks performing dense, continuous computation. Spiking neural networks process information as discrete events in time, which makes them a natural fit for temporal signals and a candidate for efficient neuromorphic hardware. This project explored whether a Liquid State Machine could separate wanted sound from noise.
What I built
A Liquid State Machine for audio denoising. Incoming sound is converted into spike trains that drive a recurrent reservoir of spiking neurons, the ‘liquid’. Because the reservoir’s activity at any moment reflects the recent history of its input, it holds a rich, time-dependent representation of the signal without the reservoir itself being trained.
A readout layer is trained on that activity to distinguish the patterns of the wanted signal from those of noise and to reconstruct a cleaner version of the sound. Only the readout learns, which keeps training lightweight and suits the event-driven, low-power computation that neuromorphic systems are designed for.
Approach
From sound to spikes and back
Key decisions
Train the readout, not the reservoir
The recurrent spiking network stays fixed and only the readout learns, which keeps training fast and stable.
Encode time, not just amplitude
Audio is converted into spike timing so the network works with how the signal changes, which is where speech and noise differ.
Present it as research
The work investigates spiking computation as a route to event-driven audio processing. It is not a production denoiser, and no benchmark figures are claimed.
Outcome
A working research prototype showing how a Liquid State Machine can learn temporal audio patterns and filter noise from sound signals. No production deployment or benchmark results are claimed; the value of the work is in the approach and what it shows about spiking networks for temporal signals.
- Capabilities
Working on something similar?
Start a conversation