Challenge
To create an exclusive and unique voice, based on the voices of real women, for Avon's virtual assistant - Ângela. In this way, she proposes to be the voice of Brazilian women to fight for more equality, more rights and more spaces, wherever they want to be.
Solution
We developed a landing page to publicize the campaign and invite women to donate their voices, storing the donated audios for future deep voice / machine learning training. All the entries were evaluated and qualified, and the best audios were selected for the training process, in this project created in partnership with Wunderman Thompson and Evil Twin.
Using the donated voices and an open dataset, we started building the deep voice. This process was divided into 3 stages: 30 hours of pre-training to arrive at a unique voice profile, construction of a base voice that was minimally neutral in terms of timbre and intonation, and finally training for approximately 300 hours to achieve the final result.
Once the training was complete, we built a dashboard with the voice generation engine, which reads any text input to transform it into Angela's original and unique voice. Based on the principle of text-to-speech (TTS), we created an exclusive brand voice for Avon.
