Voice Conversion

Voice conversion transforms speech to sound like another speaker or to have a different accent, emotion, vocal effort, or singing style while keeping the words intelligible. Common approaches disentangle linguistic content from speaker identity and prosody using self-supervised speech features, codecs, or explicit pitch representations, then generate audio with retrieval, GAN, diffusion, or flow-matching models. Many systems aim for zero-shot conversion from brief reference recordings and increasingly support streaming; evaluation commonly measures word error rate, speaker similarity, naturalness, and latency. A central challenge is balancing faithful content and target voice with natural prosody, low latency, cross-language generalization, and privacy, while detectors must also handle converted speech and other processing reliably.