NANSY
Information perturbation
Original waveform
After f perturbation
After g perturbation
Formant shift
Formant shift from 1/range(1.8~1) + range(1~1.8)
Pitch shift
Pitch shft from 0 ~ 12 (bins_per_octave=12)
Parametric Equalizer(PEQ)
Interpretable generation
Reconstructed Source Speech
Text: some people have accepted it as a miracle without physical explanation
Original Source
Reconstructed Source
Reconstructed Target Speech
Text: i cant wait to meet him
Original Target
Reconstructed Target
Converted Speech
Converted 1
Voice conversion
NANSY can perform zero-shot voice conversion by simply passing the desired target speech utterance to the speaker embedding network.
English voice conversion
The model is trained using English-only datasets (VCTK + LibriTTS-train-clean-360)Seen(Many) to Seen(Many)
Source
Converted
Target
Unseen(Any) to Seen(Many)
Unseen(Any) to Unseen(Any)
Multilingual voice conversion
Unseen Language to Unseen Language
The model is trained using English-only datasets (VCTK + LibriTTS-train-clean-360)
Source: Japanese
Converted
Target: Chinese
Source: Russian
Converted
Target: German
Source: French
Converted
Target: Spanish
Pitch shift
NANSY can perform formant-preserving pitch shift by shifting the scope of Yingram.
Semitones
Original
NANSY
Time scale modification
Ratios
Original
NANSY