this repo is a base AI SVS Model Training dataset with support files capable of the Greek language, used in the upcoming "A-MAIN -DS MODE-" Voicebank, as well as included striged samples from my dedicated Greek CVVC reclist (but thats coming in a later update tho).
Phonemes are labelled according to a custom phoneset based on X-SAMPA, with additions of [rr] and [ll], [y] replacing the default [j] and [jj], the romanized but optional [ks] and [ps] for extra coverage, as well as conventional DiffSinger-specific phonemes (or NNSVS ones).
It comes in 2 styles: X-SAMPA Style & SPHINX Style. The diffrence is that the first one retains the single capital letters for certain phonemes, and the other is entirely lowercase. you can choose whichever you like for preference.
All audio in this dataset was recorded using an AT2020USB-XP, and most samples have been treated with noise reduction, normalization, and/or a limiter. the audio's not recorded by a profesional singer (rather a local singer here), but does cover a wide male pitch and style range.
The audio data is in 16 bit WAV, 44.1 KHZ, and mono.
all hand-made .lab data files are HTK-Formatted.
By utilizing this dataset, you agree to the following terms of use:
DO'S:
- you may use the dataset's audio & labels for MultiSpeaker Training of AI SVS Models.
- you may use the dataset's audio & labels for research purposes.
- you may use the dataset's labels when recording audio for your own AI SVS Model.
DONT'S:
- you may not use the audio & labels for MultiSpeaker Training of AI SVS Models that include celebrity voices, or any other datasets that you don't have permission to use/train.
- you may not use the dataset's audio for RVC / SVC Making. distribution of the resulting model will be forbidden.
anything outside of the above usage falls under the following license with the following exceptions:
the Open-Greek-SinginDataset by A-MAIN is licensed under CC BY-NC-SA 4.0
ShareAlike does not apply to the use of the Dataset as supllimentary material for models trained via parallel training. only for direct use of models mainly featuring the voice.