Voice conversion samples
Samples from a voice-conversion system evaluated across three conversion tasks, each converting one attribute of a recording while keeping the spoken content fixed. For every example: Source is the original recording, Target reference is a real recording carrying the target attribute, and Generated is the model’s conversion of the source content into that target. The italic line above each row is the source sentence, which the generated clip preserves.
Voice (timbre) conversion
Converting a source speaker's voice into a different target speaker's voice, content unchanged. Source and target clips are from LibriTTS.
| Source | Target reference | Generated |
|---|---|---|
| “If I should not be in, wait for me.” | ||
| Speaker 2238 | Speaker 5126 | Speaker 5126 |
| “What is his name?” | ||
| Speaker 2238 | Speaker 5126 | Speaker 5126 |
| “In the old days she had dressed for her own sake to look pretty and be admired.” | ||
| Speaker 5126 | Speaker 5322 | Speaker 5322 |
| “'God willing I'll find a way to repay you,' he said, finishing his wine.” | ||
| Speaker 5322 | Speaker 8057 | Speaker 8057 |
| “Therefore it is necessary, by a final examination of their characteristics, to eliminate those features which are hostile to sociability.” | ||
| Speaker 8534 | Speaker 6415 | Speaker 6415 |
| “If public utility had interfered, that forest-the only one for miles around-would still be standing.” | ||
| Speaker 8534 | Speaker 8057 | Speaker 8057 |
| “So the government thinks.” | ||
| Speaker 8534 | Speaker 6415 | Speaker 6415 |
| “There are thousands of documents, even official documents, to prove this, if necessary.” | ||
| Speaker 8534 | Speaker 2238 | Speaker 2238 |
Accent conversion
Converting a source speaker's regional accent into a different target accent, content unchanged. Source and target clips are VCTK recordings.
| Source | Target reference | Generated |
|---|---|---|
| “She can scoop these things into three red bags, and we will go meet her Wednesday at the train station.” | ||
| English | Irish | Irish |
| “When a man looks for something beyond his reach, his friends say he is looking for the pot of gold at the end of the rainbow.” | ||
| English | American | American |
| “There is a solution, she believes.” | ||
| English | Scottish | Scottish |
| “The rainbow is a division of white light into many beautiful colors.” | ||
| Irish | English | English |
| “If the red of the second bow falls upon the green of the first, the result is to give a bow with an abnormally wide yellow band, since red and green light when mixed form yellow.” | ||
| Irish | Scottish | Scottish |
| “The difference in the rainbow depends considerably upon the size of the drops, and the width of the colored band increases as the size of the drops increases.” | ||
| American | Scottish | Scottish |
| “It is a simple equation.” | ||
| Scottish | English | English |
| “When a man looks for something beyond his reach, his friends say he is looking for the pot of gold at the end of the rainbow.” | ||
| Scottish | Irish | Irish |
Style conversion
Converting a source speaking style (happy, sad, laughing, whisper, enunciated, default) into a different target style, content unchanged. Source and target clips are from Expresso.
| Source | Target reference | Generated |
|---|---|---|
| “So, what was he really doing?” | ||
| Sad | Whisper | Whisper |
| “With exercising our first amendment rights?” | ||
| Happy | Sad | Sad |
| “With exercising our first amendment rights?” | ||
| Sad | Enunciated | Enunciated |
| “Yes, I have the doctor's appointment and checking the coffee on my to do list.” | ||
| Happy | Default | Default |
| “I was in Australia actually.” | ||
| Laughing | Default | Default |
| “We did a movie called Desperados for Netflix last year.” | ||
| Enunciated | Laughing | Laughing |
| “I'm nineteen and this is nineteen ninety eight.” | ||
| Sad | Laughing | Laughing |
| “The Norsemen considered the rainbow as a bridge over which the gods passed from earth to their home in the sky.” | ||
| Sad | Enunciated | Enunciated |
Part of ongoing doctoral work on domain-robust voice conversion. See the project overview, or the cross-speaker pair corpora used to train the underlying diffusion voice-conversion model.