On-Device Speech Enhancement
We strip noise and echo from your microphone array input and hand clean audio to the speech recognition engine you already use. No cloud. No dedicated NPU.
Built for product teams putting a voice interface into robots, kiosks, vehicles, and home appliances.
Measured in an accredited TTA test report · 65 dB speech · 60 dB noise
Because the signal reaching the microphone already carries noise, reverberation, and echo. The recognition engine receives a corrupted signal, and swapping engines does not change what arrives before them.
Nearby conversation, TV and audio playback, road noise, machinery, room reverberation — in real-world noise and reverberation, the input quality delivered to the recognition engine drops sharply.
Multiple voices arrive at the same time.
Sound played by the device loops back into the microphone.
Engine and driving noise interfere with voice commands.
The product's own operating sound mixes into the input.
The farther the user is from the microphone, the more noise dominates.
Reflections from walls and ceilings degrade recognition.
What matters more than the recognition engine is the quality of the signal it receives.
By TTA, an accredited test laboratory. The figures below are measured in a TTA test report (2023) from the Ministry of SMEs and Startups Didimdol programme, with mpAB as the article under test.
mpWAV validates before-and-after differences against real data and real product environments.
| Metric | Measured (mpWAV) | Source and test conditions |
|---|---|---|
| SRAM footprint | 0.571 MB | TTA test report (2023), Ministry of SMEs and Startups Didimdol · mpAB · static spec |
| Clock frequency | 64 MHz | TTA test report (2023), Ministry of SMEs and Startups Didimdol · mpAB · static spec |
| Real-Time Factor (RTF) | 0.852 | TTA test report (2023), Ministry of SMEs and Startups Didimdol · mpAB · test conditions not stated in the report |
| Stereo echo and noise suppression | 42.088 dB | TTA test report (2023), Ministry of SMEs and Startups Didimdol · mpAB · pass criterion 40 dB or higher |
Measured in a TTA test report (2023, Ministry of SMEs and Startups Didimdol programme) with mpAB as the article under test.
Robots, kiosks, vehicles, factories, and care settings. These are summaries — the cases we have written consent for are published in full.
Robot voice interface preprocessing
Validated preprocessing to improve a home robot's speech recognition amid TV audio, conversation, and household noise.
Speech recognition in showroom noise
Evaluated the voice interface of a showroom guide robot amid crowd conversation and room reverberation.
Care robot voice interface
We supply the voice interface module that supports voice interaction between older adults and a care robot.
Microphone-array kiosk module
Applied a multi-channel audio I/O and preprocessing module to achieve high recognition performance in store noise.
Production line acoustic inspection
Validated detecting abnormal motor sounds directly on the line in a noisy factory, without moving units to a separate soundproof room.
Everyday conversation support in noise
Validated the effectiveness and user satisfaction of the smart listening app at the Bangbae and Nonhyeon senior welfare centers.
Because suppressing noise also damages the speech you wanted, and recognition gets worse instead of better. Preprocessing that knows what to keep has to sit in front of the engine.
mpWAV focuses on reducing noise and echo while preserving the target voice recognition depends on.
Without relying on relative microphone position data, mpWAV optimizes from the actual input signals.
That reduces the repeated tuning burden when microphone count or placement changes.
You can keep the recognition engine you use today exactly as it is — no retuning for the preprocessing.
mpWAV preprocessing sits in front of it and improves the input signal quality.
Beyond algorithms, we support multi-channel audio I/O, DSP and FPGA porting, board design, and expansion to a dedicated SoC.
It removes echo, suppresses noise, and keeps only the speaker you need. Acoustic echo cancellation, noise suppression, and beamforming run before the recognition engine sees the signal.
Speech Enhancement
Multi-channel acoustic echo cancellation
Removes echo generated by the device itself — TV, car audio, voice prompts — that loops back into the microphone.
Speech Enhancement
Multi-microphone ambient noise reduction
Uses multiple microphone inputs to reduce ambient noise and strengthen the target voice.
Speech Enhancement
Integrated speech preprocessing
Integrates mpAEC and mpBeamforming to process speech in real time where noise and echo coexist.
Speech Enhancement
Single-microphone noise removal
Removes ambient noise using just one microphone, for earbuds and small smart devices where an array is not an option.
Speaker separation, diarization, wake-word detection, on-device recognition, on-device LLM, sound-source localization, and multi-channel hardware design all come after preprocessing.
Anywhere noise gets in the way. We have run PoCs in robotics, mobility, kiosks, and home appliances.
Helps service, home, care, and guide robots recognize user commands reliably amid ambient noise.
Voice interfaces for barrier-free kiosks in stores, hospitals, transit, and public institutions.
Delivers reliable recognition of the driver's voice where car audio, driving noise, and passenger conversation coexist.
Helps TVs, appliances, smart home and IoT devices work reliably in real homes.
Analyzes acoustic signals so abnormal motor and equipment sounds are detected immediately, even on noisy production lines.
Delivers dependable voice interfaces for the equipment and systems used in battlefield and armored environments with extreme noise.
Separates speech and distinguishes speakers to support meeting minutes and clear communication across meetings, voice chat, and telecom.
ClearSense Audio raises speech clarity through ordinary earphones and a smartphone, supporting smoother everyday communication.
See ClearSense AudioAs a library, a module, a port, a license, or SoC IP. Which one fits depends on where your product is in development.
Speech preprocessing algorithms delivered as a library.
A module integrating the preprocessing software with multi-channel audio I/O hardware.
We port the technology to DSP, FPGA, or AP to meet your compute and latency requirements.
License speech preprocessing or acoustic anomaly detection for your products and services.
Dedicated chips and semiconductor IP collaboration for volume production and product expansion.
With a PoC on your own data. Environment analysis, data review, PoC, then product integration — and you can start without field recordings.
We review the product, the microphone array and multi-channel audio I/O hardware configuration, user distance, noise environment, and current errors.
We review your real speech data, or a jointly designed test environment.
We compare recognition accuracy, latency, and compute before and after preprocessing.
We apply the technology as software, a module, or a DSP/FPGA architecture.
We verify performance where the product is actually used and fine-tune.
We expand to follow-up models, product lines, regions, and languages.
Smart Listening App
It refines the sound your phone picks up with AI, so ordinary earphones carry conversation clearly through noise.
Built under a Seoul Business Agency assistive-technology program, it was trialed with 100 participants at the Nonhyeon and Bangbae senior welfare centers, scoring 5.8 out of 7 for satisfaction.
FAQ
It is the stage that removes noise and echo before the signal reaches your speech recognition engine.
It raises the quality of the input signal without changing the recognition engine itself, so it applies to a speech system you have already built. mpWAV delivers echo cancellation, beamforming, and noise suppression as a single on-device module.
No. You keep the engine you have.
mpWAV is a preprocessing layer that sits between the microphone input and the recognition engine. It applies the same way to a commercial API or an in-house engine, with no retraining.
It runs in real time at 64 MHz clock, 0.571 MB of SRAM, and a 543 KB binary.
The real-time factor is 0.852, and it needs no cloud connection and no dedicated NPU. It drops into embedded environments such as robots, kiosks, and vehicles.
It was verified in an accredited third-party test.
A TTA test report from the SMBA Didimdol program recorded 42.088 dB of noise reduction and a real-time factor of 0.852 (65 dB speech, 60 dB noise). mpWAV also holds two CES 2024 Innovation Awards and NET new-technology certification from the Ministry of Trade, Industry and Energy.
You can choose a software library, a multi-channel audio I/O preprocessing module, a DSP or FPGA port, a technology license, or an SoC partnership.
Which one fits depends on your development stage and compute platform.
It starts with a PoC on your own recorded speech data.
The sequence is environment analysis, data review, PoC and performance validation, product integration, on-site tuning, and then scaling to production. We can begin with the environment analysis even if you have no field recordings yet.
Tell us your product type, audio I/O configuration, usage environment, and current errors — we will recommend the right validation method and integration structure.
You don't have to commit to a large project up front. Start a PoC with your most important environment and data.
