Multi-Channel Speech Privacy Processing for ASR Accuracy

The patent divides the speech processing into separate channels: one channel processes the original speech signal for ASR accuracy, while another channel applies voice style transfer for privacy protection. This segmentation allows each channel to optimize for its specific function without compromising the other.

The patent introduces an intermediary multi-channel processing system that receives the original speech signal and produces multiple output channels. This intermediary system enables the speech content to be preserved for ASR while simultaneously generating a privacy-protected version through voice conversion.

The patent segments the speech processing system into multiple independent channels, where one channel maintains the original speech quality for processing tasks while another channel applies voice conversion for anonymity. This allows speech processing quality to be preserved in the original channel while achieving anonymity in the converted channel.

The patent adds a dimensional aspect to speech processing by creating parallel processing channels. Instead of modifying the original signal in place, the system processes the speech in multiple dimensions (original and converted channels), allowing quality preservation in one dimension while achieving anonymity in another.

The patent implements segmentation by creating separate processing channels for privacy protection and speech recognition. The original speech channel maintains high ASR accuracy while the converted speech channel provides privacy protection, thereby reducing Word Error Rate in the original channel while achieving privacy goals.

The patent applies voice conversion selectively rather than universally. By using partial action (applying conversion only where needed for privacy while preserving original signals for ASR), the system achieves privacy protection without excessively degrading speech recognition accuracy.