Decoding Thai Phonology for AI: Teaching Text-to-Speech Models Through the Lens of Traditional Bases and Tone Classes 3

From Script to Pitch: Demystifying the Four Tone Markers and Their Mathematical Behavioral Shifts for TTS.

Table Examaple Thai TTS WannaYuke Read and Text Form

Introduction: The Misconception of Four Markers

In many Western languages, diacritics represent a shift in vowel quality or a simple stress marker. In Thai, however, the four tone markers — Mai Ek (่), Mai Tho_o (้), Mai Trii (๊), and Mai Cattawaa (๋) — are instructions for pitch modulation.

A common pitfall when training AI models on Thai text is assuming a direct, 1-to-1 mapping between a tone marker and the resulting acoustic tone. In reality, a tone marker does not possess a fixed sound; its phonetic output is entirely governed by the initial consonant class (High, Mid, Low) and the syllable structure (Live or Dead syllable). For a Text-to-Speech (TTS) engine to achieve natural-sounding inflection, it must learn the underlying algorithmic rules of these markers rather than mapping them superficially.

The Four Markers and Their Behavioral Shifts

To teach an AI model effectively, we must break down how these markers mathematically alter the pitch trajectory based on the traditional grammar rules.

1. Mai Ek (่) : The Stabilizer of Lower Pitches

Phonetic Behavior:

When applied to Mid and High-Class consonants, Mai Ek forces the pitch into a Low Tone (a flat, sustained lower frequency).

When applied to Low-Class consonants, it acts differently, shifting the pitch into a Falling Tone (starting high, then dropping rapidly).

TTS Blueprint: The model must identify the initial consonant class first. If the token is “ค่า” (Low Class + Mai Ek), the target pitch contour must be a steep descent, not a steady low pitch.

2. Mai Tho_o (้) : The Dynamic Descender

Phonetic Behavior:

For Mid and High-Class consonants, Mai Tho_o produces a Falling Tone.

For Low-Class consonants, it pushes the pitch into a sharp High Tone (or a high-rising/glottalized pitch in central Thai dialect).

TTS Blueprint: This is where acoustic models frequently stumble. The visual marker is identical, but “ข้า” (High Class) and “ค้า” (Low Class) require entirely distinct $F_0$ (Fundamental Frequency) curve shapes in the synthesis pipeline.

3. Mai Trii (๊) & Mai Cattawaa (๋) : The Mid-Class Exclusives

Phonetic Behavior:

Mai Trii creates a distinct High Tone.

Mai Cattawaa creates a Rising Tone (starting low, then sweeping upward).

TTS Blueprint: In traditional phonology, these two markers are strictly reserved for Mid-Class consonants (and certain onomatopoeic loanwords). Because their usage is highly predictable and limited, the model can easily map them to consistent high and rising pitch contours, making them the most straightforward markers for an AI to learn.

The Computational Formula for Tone Resolution

For data pre-processing or creating a custom Grapheme-to-Phoneme (G2P) conversion script, the relationship can be visualized as a systematic function:

Acoustic Tone = f(Consonant Class, Syllable Ending, Tone Marker)

Without hardcoding this linguistic logic into the tokenization layer or feeding the model a deeply balanced phonetic dataset, the AI will struggle with homographs and tone shifting, resulting in a synthetic voice that sounds unnaturally robotic or flat to a native speaker’s ear.