Introduction
In natural human communication, silence is as vital as sound. Pauses serve multiple functions: they mark syntactic boundaries, provide cognitive processing time for the listener, and convey rhetorical emphasis. However, many established text-to-speech (TTS) frameworks, such as Festival or MaryTTS, traditionally assign static, hard-coded durations (e.g., 150ms) to phrase breaks regardless of context or genre. This "one-size-fits-all" approach contributes significantly to the "robotic" quality often associated with synthetic voices. This paper proposes a data-driven methodology to capture and replicate the rhythmic "fingerprints" of diverse speaking styles.
Methodology
The researchers developed a pipeline to model pause duration across six distinct corpora. The process began with Forced Alignment using an EHMM (Expectation-Maximization Hidden Markov Model) tool to synchronize audio with text, identifying silences exceeding 80ms. Given that pause durations typically follow a non-normal distribution, a Log-Transformation was applied to the data to facilitate regression analysis.
The core of the predictive system is a CART (Classification and Regression Tree) model. These trees were trained on a variety of linguistic features extracted from the text, including:
- Part-of-Speech (POS) tags for words surrounding the break.
- Punctuation markers (e.g., commas, periods, quotation marks).
- Phrase-level metrics, such as the count of content words and stressed syllables preceding the pause.
Dataset Analysis and Style Fingerprints
The study utilized six corpora to represent a spectrum of speaking styles: Europarl (parliamentary), ARCTIC (standard TTS sentences), F2B (radio news), Obama (public addresses), and two audiobook corpora (TATS and Emma).
Analysis of these datasets revealed that each style possesses a unique pause profile. For instance, the "Obama" corpus exhibited an average pause length of 414ms, nearly triple that of the "ARCTIC" corpus (130ms). Kernel density plots of these distributions confirm that "style" is intrinsically linked to the distribution of silences.
Results and Discussion
Objective evaluations showed that style-specific CART models consistently outperformed fixed-duration baselines in correlating with natural speech patterns. The improvement was particularly notable in larger datasets like the "Emma" audiobook corpus, where linguistic variety is high.
Subjective listening tests conducted via Amazon Mechanical Turk further validated these findings. When asked to choose between speech synthesized using a style-specific CART model and a "Mean" model (which predicts the average pause length for that style), listeners showed a significant preference for the style-specific approach. While the "Mean" model often achieves lower statistical error (RMSE), it lacks the natural variability that human listeners expect.
Interestingly, when comparing style-specific models to a "combined" model trained on all six datasets, listeners could not consistently distinguish between them. This suggests that while moving away from fixed durations is critical, a sufficiently large and diverse general model may capture enough stylistic variance to be effective across multiple domains.
Conclusion
Modeling pause duration is a critical step toward more expressive and human-like speech synthesis. By moving from hard-coded constants to data-driven, style-aware models, TTS systems can better replicate the prosodic nuances of different genres. Future work will focus on modeling pauses between sentences and paragraphs to improve the flow of long-form content like audiobooks.
Summary updated