Neural-network-based noise cancellation

• Consumer IoT
• Automotive
• Mobile
• PC 

Ceva-ClearVox ENC implements a trained neural network in embedded software, eliminating the need for external cloud connectivity. Despite the neural-network approach, ClearVox ENC requires so little memory and processing power that it can be coresident with other applications even in tiny hardware configurations. Yet the algorithm is effective enough to give excellent speech quality for videoconferencing and gaming.

Working with a single microphone input for outgoing speech and—uniquely—also processing incoming signals from the other parties in a connection, ClearVox ENC effectively separates speech from both continuous and transient background noise. Versions are available for Ceva-BX1, BX2, and SensPro2 DSP cores and for ARM MCUs.

Using a trained, optimized, and highly compact neural network, ClearVox ENC effectively separates clear voice signals from both continuous and transient ambient noise, yielding intelligible speech even with adverse environmental noise. The ClearVox ENC algorithm requires only a single local microphone, and uniquely processes both outgoing and incoming speech. Thus both parties in a conference or gaming environment benefit from the ENC.

The software requires minimal memory space and computing power, so it can be run effectively even in quite small systems, and in conjunction with other applications, such as Ceva-RealSpace.

• Embedded, neural-network-based algorithm requires only a single microphone.
• Local processing means no Internet latency or cloud-AI privacy issues.
• Bidirectional processing cancels noise on both outgoing and incoming signals.
• 16 kHz sampling rate and 16-bit resolution ensure high speech quality and cancellation of broad-spectrum and high-frequency environmental noise.
• Optimized code requires little memory space or processing power, allowing use in small systems even in conjunction with other applications.