Controllable Emphasis with zero data for text-to-speech

Arnaud Joly, Marco Nicolis, Ekaterina Peterova, Alessandro Lombardi, Ammar Abbas, Arent van Korlaar, Aman Hussain, Parul Sharma, Alexis Moinet, Mateusz Lajszczak, Penny Karanasou, Antonio Bonafonte, Thomas Drugman, Elena Sokolova

Abstract: We present a scalable method to produce high quality emphasis for text-to-speech (TTS) that does not require recordings orannotations. Many TTS models include a phoneme durationmodel. A simple but effective method to achieve emphasizedspeech consists in increasing the predicted duration of the emphasised word. We show that this is significantly better thanspectrogram modification techniques improving naturalness by7.3% and correct testers’ identification of the emphasized wordin a sentence by 40% on a reference female en-US voice. Weshow that this technique significantly closes the gap to methodsthat require explicit recordings. The method proved to be scalable and preferred in all four languages tested (English, Spanish, Italian, German), for different voices and multiple speakingstyles.