How Rule-Based TTS Handled Coarticulation

Historical rule-based text-to-speech (TTS) systems, primarily utilizing formant synthesis, faced the fundamental challenge of turning discrete symbolic text into continuous, natural-sounding speech. In human speech, adjacent phonemes blend together because the physical articulators—such as the tongue, lips, and jaw—move continuously from one position to the next, a phenomenon known as coarticulation. To simulate this digitally without relying on large pre-recorded audio databases, early systems like the Klatt synthesizer and MITalk relied on acoustic locus theory, context-sensitive rewrite rules, and mathematical smoothing functions to dynamically compute formant trajectories between sounds.

The Challenge of Coarticulation in Early Synthesis

In pure phonetic transcription, phonemes are represented as isolated, static units. However, human vocal production cannot instantaneously jump from the acoustic profile of a consonant to that of a vowel. If a synthesizer plays static phoneme targets sequentially, the result is robotic, disjointed, and largely unintelligible.

To overcome this, rule-based systems had to calculate continuous changes in acoustic parameters—such as the fundamental frequency (\(F_0\)), formant frequencies (\(F_1, F_2, F_3\)), bandwidths, and nasal/fricative resonances—over time.

Locus Theory and Formant Transitions

The primary framework used to model coarticulation between consonants and vowels was the acoustic locus theory, developed at Haskins Laboratories in the 1950s.

• Virtual Targets: Researchers observed that for many stop consonants, the second formant (\(F_2\)) transitions appear to originate from a fixed, imaginary frequency point—called the "locus"—regardless of which vowel follows. For example, alveolar stops like /d/ have an \(F_2\) locus around 1800 Hz.
• Trajectory Calculation: Instead of jumping to the vowel's steady-state formant frequency, rule engines calculated a transition curve from the consonant's locus toward the vowel's target frequency.
• Variable Extent: Because the articulators rarely reach the true locus before the vowel begins, the algorithm calculated the transition starting from an intermediate point, determined by the duration and intensity of the consonant.

Context-Dependent Phonetic Rules

Before computing acoustic trajectories, rule-based systems applied phonological rewrite rules (often based on Chomsky and Halle's The Sound Pattern of English framework). These rules selected context-specific variants of phonemes, known as allophones, depending on surrounding sounds.

Rules followed conditional logic:

• Anticipatory Coarticulation (Right-to-Left): Modifying a sound in anticipation of the upcoming sound (e.g., rounding the lips for the /s/ in "suit" because the following /u/ is rounded).
• Perseverative Coarticulation (Left-to-Right): Carrying over physical configurations from the preceding sound (e.g., nasalizing a vowel that follows a nasal consonant like /m/ or /n/).

These rules determined the target values, segment durations, and transition boundaries used by the synthesis engine.

Mathematical Smoothing and Target Interpolation

Once the engine selected the allophones and their acoustic targets, it generated continuous parameter tracks using mathematical interpolation rather than linear steps.

1. Target-and-Transition Models: Each phoneme was assigned an inherent target value and an inherent duration. If a phoneme was short, the interpolation algorithm would not allow the formant parameters enough time to reach their targets, naturally modeling "target undershoot"—a major aspect of natural fast speech.
2. Filtering and Smoothing: Rather than drawing straight lines between targets, systems like Dennis Klatt's synthesizer routed parameter steps through low-pass filters or critically damped second-order differential equations. This simulated the physical inertia of human articulators, producing smooth, S-shaped transitions between contrasting acoustic states.

The Transition to Diphone Units

While purely rule-based formant synthesis successfully established continuous transitions, crafting rules for every possible phonetic combination proved prohibitively complex. By the late 1970s and 1980s, the limits of manually tuned coarticulation rules led researchers to diphone synthesis.

Instead of generating transitions via mathematical equations, diphone systems recorded speech segments cut from the steady-state center of one phoneme to the steady-state center of the next. By shifting the coarticulatory boundary into the middle of the recorded unit, the complex transition was captured physically, marking the transition from pure rule-based synthesis to concatenative methods.