Commercial
AI and ML-based solutions
R&D. Development. Consulting.
AI and ML-based solutions
R&D. Development. Consulting.
/ about the project
An application with the voices of over 40 celebrities, mimicked by artificial intelligence.
/ task
To develop voice simulation technology, overlay impersonation functionality on it. Package it into a mobile application and release it in a short period of time.
/ Solution
The neural network was initially trained on a huge amount of data (more than 30 hours of speech recordings) of one particular speaker who is not a famous person. Further training on the data of the selected celebrity lasts no more than 2 hours. In essence, the neural network has "its" voice and overlays the intonations and speech features of the person being parodied on top of it.
Synthesis technologies are packaged in a mobile application based on React Native.
In 3 months we built an MVP and managed to occupy a new niche. Saved 50% development budget by choosing cross-platform technology and speed. At the moment the application has been downloaded by more than 9 million users.
Synthesis technologies are packaged in a mobile application based on React Native.
In 3 months we built an MVP and managed to occupy a new niche. Saved 50% development budget by choosing cross-platform technology and speed. At the moment the application has been downloaded by more than 9 million users.
/ technologies
Python, React
NLP frontend, which includes neural network text normalization and a model for pause and accent placement.
Tacotron 2, which takes letters as input.
Autoregressive WaveNet running in real time on the CPU.
NLP frontend, which includes neural network text normalization and a model for pause and accent placement.
Tacotron 2, which takes letters as input.
Autoregressive WaveNet running in real time on the CPU.