Development of speech enhancement algorithms for directional microphones - QC-268Desired discipline(s): Engineering - computer / electrical, Engineering, Computer science, Mathematical Sciences, Mathematics
Project Length: 6 months to 1 year
Preferred start date: 05/01/2020
Language requirement: Flexible
Location(s): Montreal, QC, Canada
No. of positions: 1
About the company:
Soundskrit is an audio start-up based in Montreal. We are developing the world’s first multi-directional microphone chip and the associated algorithms to improve sound capture for consumer electronics applications.
The many emerging voice applications and interfaces have driven a need to capture sound from farther distances while remaining robust to noise. Because traditional pressure-based microphones pick up sound coming from every direction, arrays of multiple microphones are needed to listen into a specific direction and ignore surrounding noise. This comes at a cost of increased size and decreased audio quality.
Our bio-inspired, directional microphone directly measures the velocity of incoming sound rather than pressure. With a single microphone, we can separate sounds coming from multiple directions without compromising on size or quality. Soundskrit will greatly improve speech recognition and sound localization, enabling a new suite of features to consumer electronics
Please describe the project.:
Using the analog outputs of our new sensor, we are developing an algorithmic solution that can extract a single channel of clean speech signal, even when the recordings are corrupted by room reverberation or mixed with other undesired sounds in the background.
This cleaned signal will enable much better performance for voice recognition service (Alexa, Google) that cannot properly deal with noise otherwise. In phone call applications, our solution maximizes speech quality and speech intellibility for a better user experience.
This project aims at improving our audio chain and designing better algorithms to push the performace of our solution furter: being able to listen to a speaker further away by reducing the distortion of our algorithms and increasing the robustness to all kinds of noise.
Main objectives of the project:
- Conduct a review our current audio chain and identication the weaker algorithms
- Recommend better approaches based on the state of the art for microphone arrays algorithms (multi-channel noise reduction, de-reverberation, sound localization and beamforming)
- Implementation of the recommended algorithm and writing associated documentation
- Establish theorerical limits for algorithm performance using acoustic models an real recordings of corner cases
- Recommendation of follow-up work and research effort to go beyond the state-of-the-art (ex : use of machine learning or other)
- Experience with algorithms for speech emhancement with focus on multi-channel aspects: noise reduction, acoustic echo cancellation, beamforming etc.
- Knowledge of modern audio systems (conference phone, smart speakers)
- Knowledge of Matlab or Python
- Knowledge of acoustics and psycho-acoustics