Continuous Pitch vs Discontinuous Tracks in TTS

Continuous pitch representations provide substantial advantages over discontinuous pitch tracks in Text-to-Speech (TTS) synthesis by eliminating unvoiced gaps, stabilizing neural network training, and improving prosodic naturalness. Traditional fundamental frequency (\(F_0\)) representations define pitch only during voiced segments, creating discontinuous step changes and undefined values during unvoiced consonants and pauses. Transitioning to a continuous pitch contour—typically paired with an independent voicing predictor—decouples global intonation from segmental phonetics, resulting in smoother gradient propagation, fewer acoustic artifacts, and more human-like speech output.

Elimination of Discontinuous Gradient Boundaries

Standard discontinuous pitch contours assign a zero, negative, or undefined value to unvoiced segments. When training deep learning acoustic models or vocoders, these sudden transitions between voiced values (e.g., 100–300 Hz) and zero values act as sharp mathematical step functions.

Continuous representations interpolate pitch values through unvoiced regions, generating a continuous trajectory. This continuous curve enables standard regression losses (such as Mean Squared Error or L1 loss) to propagate gradients smoothly across the entire utterance without destabilizing updates at voiced-to-unvoiced transitions.

Decoupling Intonation from Segmental Voicing

Linguistic intonation exists as a continuous cognitive and physiological target across an entire utterance, even when unvoiced consonants (such as /s/, /t/, or /k/) interrupt physical vocal cord vibration. Discontinuous pitch representations conflate two separate physical mechanisms:

• Phonation: Whether the vocal folds are vibrating (voiced vs. unvoiced).
• Pitch contour: The tension and frequency of the vocal folds determined by the broader prosodic contour.

Continuous pitch representations separate these phenomena. The continuous \(F_0\) model predicts the underlying macro-prosody and emotional contour, while an auxiliary binary classifier predicts the voicing state. This separation allows the system to capture long-range melodic patterns without being disrupted by local phonetic obstacles.

Improved Model Convergence and Stability

In neural TTS architectures (such as FastSpeech, VITS, or Tacotron variants), discontinuous pitch inputs force models to learn complex non-linear mappings to ignore unvoiced regions. This often results in:

• Slower convergence rates during the training phase.
• Training instability and sensitivity to weight initialization.
• Pitch "leakage" where models predict low, non-zero frequencies into unvoiced regions, causing audible buzzes or raspiness.

Continuous representations present a smoother distribution for the model to fit. By normalizing and modeling continuous trajectories, statistical and neural predictors converge faster and achieve lower variance in their objective evaluation metrics.

Mitigation of Boundary Artifacts

Vocoders that depend on discontinuous \(F_0\) contours often struggle at the onset and offset of voicing. Slight timing misalignments between the predicted pitch curve and the predicted spectral features can lead to:

• Harsh clicking sounds at the start of voiced phonemes.
• Unintended micro-prosodic dips or spikes at vowel-consonant boundaries.
• Robotic transitions caused by sudden zero-value dropouts.

Using a continuous pitch track ensures that the vocoder always has an established frequency target at the exact moment voicing starts. When the voicing decision flips to active, the vocoder synthesizes the correct pitch immediately, avoiding phase inconsistencies and onset glitches.

Enhanced Expressiveness and Contour Controllability

Expressive TTS systems often require pitch scaling, shifting, or style transfer. Manipulating a discontinuous track is difficult because scaling factors can inadvertently alter unvoiced thresholds or create unnatural boundary cliffs. Continuous pitch representations can be modified, smoothed, or transformed using techniques like continuous wavelet transforms (CWT) or spline adjustments, allowing fine-grained, intuitive control over prosody without distorting phoneme-level boundaries.