No abstract

No abstract

Pause-internal phonetic particles (PINTs), such as breath noises, tongue clicks and hesitations, play an important role in speech perception but are rarely modeled in speech synthesis. We developed two text-to-speech (TTS) systems: one with and one without PINTs labels in the training data. Both models produced fewer PINTs and had a lower total PINTs duration than natural speech. The labeled model generated more PINTs and longer total PINTs durations than the model without labels. In a listening experiment based on the labeled model we evaluated the influence of various PINTs combinations on the perception of speaker certainty. We tested a condition without PINTs material and three conditions that included PINTs. The condition without PINTs was perceived as significantly more certain than the PINTs conditions, suggesting that we can modify how certain TTS is perceived by including PINTs.

For isolated utterances, speech synthesis quality has improved immensely thanks to the use of sequence-to-sequence models. However, these models are generally trained on read speech and fail to generalise to unseen speaking styles. Recently, more research is focused on the synthesis of expressive and conversational speech. Conversational speech contains many prosodic phenomena that are not present in read speech. We would like to learn these prosodic patterns from data, but unfortunately, many large conversational corpora are unsuitable for speech synthesis due to low audio quality. We investigate whether a data mixing strategy can improve conversational prosody for a target voice based on monologue data from audiobooks by adding real conversational data from podcasts. We filter the podcast data to create a set of 26k question and answer pairs. We evaluate two FastPitch models: one trained on 20 hours of monologue speech from a single speaker, and another trained on 5 hours of monologue speech from that speaker plus 15 hours of questions and answers spoken by nearly 15k speakers. Results from three listening tests show that the second model generates more preferred question prosody.