声のクローンと変換
わずかなサンプルから声質を再現する
本ページの本文は英語で提供されています。タイトルと導入は日本語化されています。
この能力とは何か
Takes a reference clip (the target timbre) and the text to be spoken, and outputs that text in that voice. It splits into two kinds: zero-shot cloning needs only a few seconds of audio, while voice conversion keeps the content and timing of the source and merely swaps the timbre. Unlike speech synthesis the timbre comes from a sample rather than a preset.
技術的にどう実現するか
The speaker timbre is encoded into a vector disentangled from content and injected as a condition into the acoustic model; voice conversion substitutes the target timbre features into the source audio and rebuilds the waveform. Zero-shot systems train on large multi-speaker data so the timbre encoder generalises to unseen speakers. For compliance they are usually paired with consent checks, watermarking and synthetic-audio markers to limit misuse.
代表的な製品
3関連する組織
代表的な用途
- A consistent narrator voice across content
- Keeping the original voice in multilingual dubbing
- Accessibility and voice restoration for those who lost speech
- Character voices in games and animation
どう評価するか
- Speaker similarity (SECS)
- Cosine similarity between output and target speaker embeddings
- MOS
- Mean opinion score for naturalness
- Equal error rate (EER)
- Error rate where false accept and false reject meet, gauging impersonation risk
限界と難しさ
- With only seconds of reference audio the timbre is unstable, sounding like two speakers across sentences
- Cross-lingual or cross-emotion transfer drifts, so timbre shifts with the language
- It can be used to impersonate people, so consent and watermarking are required or the social-engineering and fraud risk is high