Acoustic measurement & calibration
Characterize microphones, sensors and the measurement chain. Examine sensitivity, frequency response, noise floor and dynamic range to understand what the data can support.
MLAIA Audio & Acoustics
Intelligence starts
with the signal.
Acoustics, classical signal processing and machine learning, engineered together. From microphones and vibration sensors to algorithms running on real devices.
A specialist practice of MLAIA Data ScienceLed by Dr. Yochai Edlitz · Ph.D., Weizmann Institute
Classical acoustics & digital signal processing
Understand the physics.
Engineer the signal.
A useful audio system starts with how sound is generated, captured and transformed. We work across the measurement setup, signal chain and algorithm—so model decisions are grounded in what the sensors actually measure.
Time & frequency analysis
Use FFTs, spectrograms, spectral estimates and correlation to investigate harmonics, transients, periodicity and interference. Connect observed patterns to physical sources.
Filter design & noise reduction
Design frequency-selective filters and assess noise-reduction strategies. Evaluate the tradeoff between unwanted-signal suppression, useful-signal preservation and processing delay.
Hear the filtering experiment →Microphone arrays & beamforming
Use spatial information to focus on a direction. Account for array geometry, propagation delays, frequency-dependent response and the acoustic environment.
Steer the array simulation →Vibration & physical sensing
Analyze vibration signals, resonances and sensor responses. Design repeatable experiments and interpret measurements in the context of the underlying mechanics.
Real-time signal pipelines
Engineer sampling, buffering, filtering and feature extraction around the target device. Evaluate the complete path from acquisition to output, including latency and compute limits.
Choose the method that fits the problem
Classical DSP. Machine learning.
Or a carefully designed combination.
Start with a signal-processing baseline. Add learned models when the task calls for them—for example, acoustic event detection or speech recognition. A combined system can use DSP to condition the signal and a model to interpret it, with each stage evaluated against the same operating requirements.
For your product, that means testing accuracy alongside signal quality, robustness, latency, memory and power—not treating the model score as the whole system.
Explore the audio lab →Semantic hearing & audio intelligence
Understand which sounds matter.
Move from measuring a signal to interpreting its meaning for an application. Combine acoustics, signal processing and learned models to recognize relevant events and support selective listening.
Sound-event recognition
Define the sounds a system needs to recognize, collect representative examples and evaluate detections against false alarms and missed events.
Listening for a purpose
Connect sound classes and context to a user’s needs. Decide which events to preserve, emphasize or act on, and evaluate the whole listening experience.
Intelligence on the device
Balance recognition quality, latency, memory and power on phones and embedded hardware. Test with the noise, microphones and operating conditions the product will encounter.
The current interactive lab illustrates classical signal-processing principles with synthetic audio. It does not run a semantic sound-recognition model.
Specialist topics
Go deeper on
a specific problem.
Each topic page explains how we approach the problem, what we measure and the questions to settle before a project starts.
Microphone arrays & beamforming
Array geometry, direction-of-arrival estimation and fixed or adaptive beamformers, designed around the device and the room.
Beamforming & arrays →Noise reduction & ANC
Speech enhancement, echo cancellation and active noise cancellation, balancing suppression against distortion and delay.
Noise reduction & ANC →On-device audio AI
Wake words, sound events and speech models on phones, DSPs and dedicated controllers, within memory, latency and power budgets.
On-device audio AI →Vibration analysis
Accelerometer and vibration signals for condition monitoring and product sensing, from sensor mounting to fault features and anomaly detection.
Vibration analysis →Client perspective
Recognized for audio and DSP expertise.
“Your expertise and vast knowledge have been instrumental in establishing our AI infrastructure.”
Tzvika Fridman — CTO, Silentium
Excerpt from the testimonial published on MLAIA’s website.
Voice interfaces
Speech that connects to your workflow.
Speech-to-text (STT), text-to-speech (TTS) and domain-specific audio models. Evaluate language, vocabulary, noise and latency together, then connect the voice interface to an assistant or business process.
Edge, embedded & physical AI
From a trained model
to a working device.
Our project experience includes deploying machine-learning and deep-learning models on Android and iOS phones, as well as dedicated controllers for applications involving drones and earphones.
Mobile inference
Run models inside Android and iOS applications. Design around device resources, sensor access, responsiveness and the experience of the person using the app.
Dedicated controllers
Bring inference to constrained hardware. Assess model size, memory, processing time and power alongside the application’s accuracy requirements.
AI in physical systems
Connect sensing and model inference to systems operating in the physical world. Define how predictions inform device behavior, with operating limits and validation on the target hardware.
Engineering considerationsModel compression & quantizationSignal pipelinesLatency & power budgetsDevice-level evaluation
How we work
Measure. Model. Validate on the device.
Frame the problem
Agree on the decision, available data, operating constraints and success criteria.
Establish a baseline
Characterize the signal and measurement chain, design repeatable tests and establish a classical DSP baseline.
Develop & evaluate
Compare DSP, learned and combined approaches where relevant. Review failure modes and tradeoffs using representative signals.
Integrate & hand over
Plan deployment, monitoring, documentation and ownership around the agreed scope.
From our knowledge base
Explore the thinking behind the work.
Before we start
Good questions. Straight answers.
Does every audio problem need a neural network?
No. Filtering, spectral analysis or a physical model may be sufficient for a well-defined task. We select the approach using measured performance and the product’s constraints. Where machine learning adds value, we evaluate it as part of the full signal chain.
Can you improve an existing audio model?
Yes. We can start with a focused review of the data, labeling, signal chain and deployment constraints, then agree on an evaluation plan before changing the model.
Can you guarantee a particular latency or accuracy?
Targets must be established against your hardware, operating environment and representative test data. We define the full latency budget and relevant error measures with you; we do not promise untested benchmark numbers.
Do you work beyond speech?
Yes. The practice includes acoustic events, vibration and sensor signals, combining physical understanding with statistical and machine-learning methods.
Audio & Acoustics
Tell us what
you’re working on.
Share the problem, the data you have and what success would look like. We’ll discuss a practical next step.
Please don’t include confidential datasets, credentials or patient information.