Giving synthetic speech room to breathe

01 / MOTIVATION

The question

How can a text-to-speech system learn where to pause, and how do those pauses change the listening experience?

02 / METHOD

The approach

The early work investigates word-terminal syllables for phrase break prediction in Indian languages (SSW 2013), followed by continuous-valued word representations (Interspeech 2014) and recurrent neural architectures with word embeddings (Interspeech 2016).

The later work integrates predicted phrase breaks into a Tacotron2–WaveRNN pipeline with dynamic convolutional attention. BERT is fine-tuned for phrase break prediction on LibriTTS, and synthesized stories are assessed in a 70-listener study.

03 / EVIDENCE

Findings & scope

The later phrase prediction model achieved 92.10 F1, compared with 88.91 for a BiLSTM baseline. Pairwise listening tests favored speech with predicted phrase breaks over synthesis without phrase break prediction.

Prediction scores and listener preference answer different questions; both inform the evaluation. This thread underpins the Ph.D. thesis on prediction and incorporation of phrase break models in TTS systems.