Neural TTS for Kannada -Indistinguishable from human speech

State-of-the-art Deep Neural network based TTS engine developed for Kannada

RaGaVeRa's Kannada TTS exceeds the quality of Google and Nuance Kannada TTS

Comparison of the quality of RaGaVeRa’s TTS against Google’s WaveNet and Nuance’s Kannada TTS as assessed by 55 natives of Kannada. RaGaVeRa’s TTS got a mean preference score of 78.2% in contrast to 13.1% for Google’s TTS and 5.1% for Nuance’s TTS.

Our TTS system was able to achieve an MOS of 4.42 ± 0.53 when compared to the ground truth 4.55 ± 0.56 based on the evaluation of 25 Kannada natives.

A survey was conducted to compare our synthesized speech with the human speech and it was seen that people preferred the synthesized speech over human speech.

Below is a side by side comparison of the synthesized speech from RaGaVeRa, Google and Nuance's Kannada TTS:

RaGaVeRa Kannada TTS

Google Kannada TTS

Nuance Kannada TTS

01_RaGaVeRa.wav
01_Google.wav
01_Nuance.wav
02_RaGaVeRa.wav
02_Google.wav
02_Nuance.wav
03_RaGaVeRa.wav
03_Google.wav
03_Nuance.wav
04_RaGaVeRa.wav
04_Google.wav
04_Nuance.wav
05_RaGaVeRa.wav
05_Google.wav
05_Nuance.wav

Below is a side by side comparison of human speech with synthesized speech:

Human Speech

Synthesized speech

gtaudio0.wav
saudio_0.wav
gtaudio1.wav
saudio_1.wav
gtaudio2.wav
saudio_2.wav
gtaudio3.wav
saudio_3.wav
gtaudio4.wav
saudio_4.wav

Below are some more audio samples synthesized using RaGaVeRa's Kannada TTS:

01.wav
03.wav
05.wav
06.wav
08.wav
09.wav
10.wav
11.wav