|
Coming soon to VCP: Alango’s Deep Neural |
|
|
 |
|
A core focus at Alango is the separation of noise from speech, and our commitment to improvement never wavers. Classical DSP-based noise reduction, exemplified in our current-generation voice communication solutions - VCP and eVCP, stands as an effective and well-proven method for eliminating noise from speech. This technique finds utility across diverse devices and call scenarios. However, classical audio signal processing comes with its limitations. Notably, handling non-stationary noises presents challenges, and certain sounds evident to the human ear as noise, such as knocking or wind sounds, can still be misconstrued as voices in the classical approach. To enhance VCP's capability in distinguishing between noise and voice, we've been actively developing a Deep Neural Network Noise Reduction algorithm. Our aim is to seamlessly integrate this innovation into VCP by the conclusion of 2023. We are eager to share our progress, as we firmly believe it will yield significant enhancements, particularly in more intricate use cases. Below, we present audio examples that demonstrate our prototype deep learning noise reduction algorithm's performance. These include a sample from a noisy automotive cabin and another with pronounced wind noise. For a comparative evaluation, we processed the signals separately using RNNoise, a publicly available library for neural network noise reduction. Like RNNoise, Alango's prototype algorithm boasts a compact design (with under 80k weights, compared to RNNoise's 85k), swift performance, and the ability to facilitate real-time processing with moderate resource requirements. However, in contrast to RNNoise outputs, our algorithm notably reduces distortion in the user's voice.
Use control buttons on the interactive screenshots to play/stop/navigate
|
|
Automotive Cabin Noise
UNPROCESSED SIGNAL, speech recorded in car cabin
|
| Processed by Alango Deep Neural Network Noise Reduction |
| Processed by RNNoise |
|
Wind Noise
UNPROCESSED SIGNAL, high wind noise
|
| Processed by Alango Deep Neural Network Noise Reduction |
| Processed by RNNoise |
|