Skip to the content.

TTS Model comparisons with an inhouse single speaker dataset

Here are audio samples synthesized from different TTS architectures. The dataset is 10 hours , 42 minutes of a single female speaker. It was recorded in a studio environment for the best output.

sentences

Ground truth

Some sample original recordings from the inhouse dataset.

Sample Text Audio
01 Okwogera mu ngeri ey’amakulu mu mbeera yonna.
02 Omu ku booluganda b’abavubuka bano, Francis Obbo bwe yagambye.
03 Amawanga agerinaanye getaaga okukwatagana.
04 Kirungi nnyo omulunzi okufuuyira embuzi ze okusobola okuzigobako enkwa.

English Hifigan v2 vocoder by coqui

Details about the model: (todo: link) Tacotron2 + DDC: 302k steps trained + Hifigan_v2 coqui vocoder

Sample Text Audio
01 Amaanyi poliisi g'etadde ku muyiggo
02 Musono ki ogwo ogw’enviiri?
03 Omuganda amanyi nti omuntu alina emmeeme bbiri.
04 Tolya wadde okubaaga ekinyonyi ekifudde oba ekirwadde.
05 Obulwadde buno businga kuba bwa mutawaana mu bifo ebinnyuukirivu ennyo, omutera okutonnya enkuba ennyingi ate ng’ebitooke bisimbiddwa kumuukumu.

WaveRNN

todo

WaveGrad

todo

HifiGAN

todo

End to End TTS architecture - VITS

VITS: 63k steps trained

Sample Text Audio
01 Amaanyi poliisi g'etadde ku muyiggo
02 Musono ki ogwo ogw’enviiri?
03 Omuganda amanyi nti omuntu alina emmeeme bbiri.
04 Tolya wadde okubaaga ekinyonyi ekifudde oba ekirwadde.
05 Obulwadde buno businga kuba bwa mutawaana mu bifo ebinnyuukirivu ennyo, omutera okutonnya enkuba ennyingi ate ng’ebitooke bisimbiddwa kumuukumu.

GlowTTS / Waveglow

todo

Speedy Speech

For better RTF ( It is the fastest model on coqui) todo

Fast Speech

todo