Epoch-Synchronous vs Fixed-Frame Processing in TTS

This article examines the core signal processing differences between epoch-synchronous and fixed-frame processing in parametric Text-to-Speech (TTS) synthesis. While fixed-frame processing divides audio into uniform, regularly spaced intervals, epoch-synchronous processing dynamically aligns analysis and synthesis windows with physical glottal events known as Glottal Closure Instants (GCIs). The following sections break down how these two paradigms diverge in windowing mechanisms, spectral estimation accuracy, phase representation, computational robustness, and overall synthesis quality.

Frame Boundary Determination and Windowing

The primary mechanical difference lies in how analysis windows are placed along the speech signal:

• Fixed-Frame Processing: The signal is segmented using a constant frame shift (typically 5 ms or 10 ms) and a fixed window length (typically 20 ms to 30 ms). The placement of the window is entirely independent of the fundamental frequency (\(F_0\)) or the state of the vocal folds.
• Epoch-Synchronous Processing: Analysis frames are tied directly to pitch epochs, specifically Glottal Closure Instants (GCIs). The frame shift is variable and equal to the local pitch period (\(T_0\)). The window length is typically set dynamically to cover a specific number of pitch periods—most commonly two fundamental periods (\(2T_0\)) centered precisely at the GCI.

Spectral Envelope Estimation and Leakage

The interaction between the analysis window and the underlying pitch period significantly impacts the accuracy of the vocal tract transfer function estimation:

• Fixed-Frame Smearing: Because the window length does not match local pitch cycles, a fixed-frame window captures an arbitrary, non-integer number of pitch pulses. This creates spectral leakage and pitch-synchronous modulation artifacts in linear predictive coding (LPC) or mel-cepstral analysis. As a result, formant bandwidths and center frequencies fluctuate artificially from frame to frame based merely on where the window lands relative to the pulse.
• Epoch-Synchronous Precision: By centering a two-period window on a GCI, the analysis window captures consistent energy profiles of the glottal cycle. This reduces pitch-dependent spectral modulation, limits cross-talk between the source harmonics and vocal tract resonances, and yields sharper, more accurate formant estimates with minimal spectral smearing.

Source-Filter Separation and Excitation Modeling

Parametric TTS vocoders rely on separating the source (glottal airflow) from the filter (vocal tract):

• Fixed-Frame Approximations: Vocoders using fixed frames (such as standard WORLD or STRAIGHT) compute a smoothed spectral envelope and model the excitation as a mix of periodic impulses and aperiodic noise bands per frame. Phase information is typically discarded or approximated via minimum-phase assumptions, which can introduce a "buzzy" or unnatural timbre.
• Epoch-Synchronous Excitation: Epoch-aligned frameworks allow direct decomposition of the speech signal into the true glottal flow derivative and vocal tract response. Models like the Liljencrants-Fant (LF) model can be accurately fitted to the signal cycle-by-cycle. Because the timing of each acoustic impulse is preserved relative to the vocal tract filter, phase alignment across pitch pulses is naturally maintained.

Phase Coherence and Artifact Generation

Phase handling differs considerably between the two methods during the synthesis stage:

• Fixed-Frame Phase Reconstruction: Because frame shifts do not match pitch periods, reconstructing the continuous phase across frame boundaries requires explicit phase unwrapping or minimum-phase plus noise-dispersion networks. Discontinuities at frame boundaries can lead to phase cancellation, roughness, or temporal blurring of transient sounds like plosives.
• Epoch-Synchronous Coherence: Pitch-Synchronous Overlap-Add (PSOLA) and epoch-based parametric vocoders naturally maintain phase alignment because synthesis grains overlap precisely at the excitation points. This eliminates destructive interference between adjacent harmonics and maintains natural waveform peakiness, reducing buzzy artifacts in voiced regions.

Computational Complexity and Robustness

While epoch-synchronous processing offers theoretical advantages in fidelity, it introduces major operational constraints:

• Fixed-Frame Robustness: Fixed-frame algorithms are computationally deterministic, simple to parallelize, and fully autonomous. They require no prior pitch marking and do not fail when acoustic conditions deteriorate.
• Epoch-Synchronous Vulnerability: Epoch-synchronous processing depends entirely on the reliability of GCI detection algorithms (e.g., SEDREAMS, DYPSA, or deep-learning-based epoch detectors). If a GCI is missed, mislocated, or falsely detected in unvoiced regions, severe spectral distortions and audible clicks occur. This dependency makes epoch-synchronous pipelines more complex and less robust when processing noisy or non-modal speech (such as vocal fry or breathy voice).

Comparison Summary