Official Archival Milestone: The 20-Year Foundational Genesis of Sodupath.
Bilingual Abstract & සිංහල සාරාංශය
English: Speech recognition in morphologically rich, low-resource languages poses profound theoretical and computational challenges. While Hidden Markov Models (HMMs) served as the predominant stochastic paradigm of the early 2000s, statistical acoustic models struggled to bridge the gap between acoustic variability and canonical orthographic representation in non-Latin scripts. In the Sinhala language—an insular Indo-Aryan language characterized by an alphasyllabic (Abugida) script, rich vowel length contrasts, and a complex consonantal inventory including prenasalized stops—phoneme-to-grapheme (P2G) transcription requires deterministic structural modeling. This monograph presents the foundational engineering and empirical validation of a digitized Phoneme-to-Grapheme conversion architecture developed at the University of Kelaniya in 2006.
සිංහල සාරාංශය: ස්වභාවික භාෂා සැකසුම (NLP) සහ කථන හඳුනාගැනීම (Speech Recognition) විෂයයෙහි ලා සිංහල භාෂාවේ ශබ්ද විද්‍යාත්මක සංඥා සහ අක්ෂර වින්‍යාසය (Orthography) අතර පවතින සබඳතාව ගණිතමය වශයෙන් ආකෘතිගත කිරීම මෙම පර්යේෂණයේ අරමුණයි. 2006 වර්ෂයේදී කැලණිය විශ්වවිද්‍යාලයේදී සම්පාදනය කරන ලද මෙම පර්යේෂණය, සිංහල භාෂාව වෙනුවෙන් සංඥා සැකසුම් (MATLAB / COLEA Signal Processing) සහ ශබ්ද-අක්ෂර පරිවර්තන (Phoneme-to-Grapheme - P2G) ආකෘතියක් ගොඩනැගූ පුරෝගාමී ප්‍රයත්නයයි. මෙහිදී කථන සිංහලයේ ශබ්ද ඒකක 40 ම (ස්වර 14 සහ සඤ්ඤක අක්ෂර ඇතුළු ව්‍යංජන 26) ආවරණය වන පරිදි ධ්වනි විස්තාරය, ශක්ති මට්ටම් (Energy levels) සහ රේඛීය සංගුණක (Y = p1 * x + p2) විශ්ලේෂණය කර දත්ත සමුදායක් ගොඩනංවන ලදී. එම ශබ්ද සංඥා ලිඛිත අක්ෂර බවට පත් කිරීම සඳහා ත්‍රිත්ව බෆරයක් [P1, P2, P3] සහිත අක්ෂර එකලස් කිරීමේ ඇල්ගොරිතමය (Phoneme Assemble Algorithm - PAA) සහ ද්විමාන අනුකෘති සිතියම්කරණයක් හඳුන්වා දෙන ලදී.
1. Introduction & Theoretical Background
Speech recognition aims to accurately map raw continuous acoustic pressure waves into meaningful written orthographic representations. In 2006, the overwhelming consensus in Automatic Speech Recognition (ASR) centered on stochastic modeling, specifically Hidden Markov Models (HMMs) coupled with Gaussian Mixture Models (GMMs) (Rabiner & Juang, 1993; Jelinek, 1997). In HMM-based systems, decoding speech is fundamentally probabilistic:
While this probabilistic formulation proved effective for English and European languages with massive statistical corpora, it encountered systemic obstacles when applied to morphologically complex, low-resource Asian languages. For English, the relation between spelling and sound is notoriously opaque, requiring extensive pronunciation dictionaries. Conversely, Sinhala (සිංහල)—an insular Indo-Aryan language spoken by over 17 million people—presents a transparent relationship between spoken phonemes and canonical graphemes in its basic phonetic realization, yet it employs an Abugida (alphasyllabary) script where vowels are not written as independent letters following consonants, but as post-, pre-, sub-, or circum-graphs (diacritics or පිලි / pili) anchored to a base consonantal glyph.
2. Phonological Architecture of Spoken Sinhala (40 Segmental Phonemes)
Spoken Sinhala (කථා සිංහල - Kathā Sinhala) possesses a distinct phonological inventory differing in several crucial dimensions from classical literary Sinhala (Karunathilake et al., 2004). The segmental inventory analyzed in this research consists of 40 segmental phonemes:
- Vowels (14 Phonemes): 7 short/long pairs: /i, iː/ (ඉ, ඊ), /e, eː/ (එ, ඒ), /æ, æː/ (ඇ, ඈ), /a, aː/ (අ, ආ), /ə, əː/ (අ), /u, uː/ (උ, ඌ), /o, oː/ (ඔ, ඕ). The low-front vowel pair /æ, æː/ is typologically rare in Indo-Aryan and developed via historical umlaut.
- Consonants (26 Phonemes): Velar (/k, g/), Palatal (/c, j/), Retroflex (/ʈ, ɖ/), Dental (/t, d/), Labial (/p, b/), Nasals (/m, n/), Liquids (/r, l/), Fricatives (/s, h, f/), and Glides (/y, v/).
- Prenasalized Voiced Stops (සඤ්ඤක අක්ෂර): /ᵑɡ/ (ඟ), /ᶯɖ/ (ඬ), /ⁿd/ (ඳ), and /ᵐb/ (ඹ). Unlike nasal-stop consonant clusters (e.g. /mb/ in කම්බි), prenasalized stops have the temporal duration of a single voiced stop, where the velum is lowered only during the initial 15–25% of the closure duration.
3. Acoustic Signal Processing & Linear Energy Regression
The experimental framework was implemented in MATLAB utilizing specialized speech signal analysis modules (COLEA). Speech was captured via high-fidelity microphones, sampled at 16 kHz with 16-bit quantization. To capture the dynamic trajectory of acoustic intensity across each phoneme segment without requiring heavy computational parameterization, we formulated a linear regression curve fitting over the decibel energy contour:
where x is the normalized temporal sample index, Y is measured energy in decibels (dB), p1 (Coefficient 1 - Slope) represents the rate of energy accretion or decay, p2 (Coefficient 2 - Intercept) represents baseline sound pressure level, and the Norm of Residuals (||r||) quantifies articulatory stability across the segment.
4. Empirical Trajectory Parameters Table
Table 1 outlines the empirical trajectory parameters obtained across representative Sinhala phonemes from native speaker recordings:
| Context | IPA | Sinhala | Train p1 | Train p2 | Norm ||r|| | Test p1 | Test p2 |
|---|---|---|---|---|---|---|---|
| පසළොස් | /a/ | අ | -0.38 | 49.54 | 4.33 | -0.44 | 49.50 |
| ආචාර්ය | /aː/ | ආ | -0.23 | 48.84 | 4.29 | -0.55 | 50.43 |
| ඇම්බල් | /æ/ | ඇ | -0.90 | 48.06 | 2.70 | -0.17 | 48.06 |
| කුඩය | /u/ | උ | -0.86 | 49.80 | 4.57 | -0.10 | 49.86 |
| කඩය | /k/ | ක | -0.59 | 46.39 | 3.21 | -0.07 | 48.49 |
| ගෙම්බා | /g/ | ග | 1.27 | 48.66 | 2.50 | -1.12 | 48.66 |
| තනතුර | /t/ | ත | 3.16 | 44.62 | 0.00 | 1.01 | 48.35 |
| දොරටුව | /d/ | ද | 0.08 | 49.21 | 0.96 | 1.07 | 48.36 |
| ළඟ | /ᵑɡ/ | ඟ | 0.61 | 51.69 | 1.54 | 0.20 | 51.69 |
| සවස | /s/ | ස | 0.47 | 49.90 | 2.72 | 0.02 | 49.90 |
5. The Phoneme Assemble Algorithm (PAA) State Machine
The Phoneme Assemble Algorithm operates over a sliding three-member buffer [P1, P2, P3] to resolve whether incoming sound units compose
independent vowel letters, consonant-vowel combinations (diacritics), or consonantal ligatures:
6. Twenty-Year Retrospective (2006–2026): From MATLAB to Sodupath
The 20-Year Evolution of Sinhala Computational Linguistics
Re-evaluating this 2006 research in 2026 demonstrates the vital importance of deterministic linguistic rules. Modern deep learning models like OpenAI Whisper or Meta MMS process audio through unconstrained neural representations, frequently hallucinating homophone substitutions (such as confusing dental න/ල with retroflex ණ/ළ) because acoustic signals alone cannot disambiguate classical literary grammar.
By uniting the 2006 acoustic-phonetic P2G pipeline with the 2026 Sodupath Deterministic NLP Engine (which enforces Sidath Sangarawa orthography and Nominative Agent Precedence), we bridge the gap between spoken acoustics and pristine literary grammar, establishing a culturally sovereign, zero-hallucination language stack.
7. References
- Hewage, V. P. (2006). Develop a Digitized Phoneme to Grapheme Conversion Model for Sinhala Language. Master's Thesis, University of Kelaniya.
- Karunathilake, W. S., Rajapaksha, R. M. W., & Wijethunga, R. (2004). Vāgvidyā Pravēśaya. S. Godage & Brothers, Colombo.
- Rabiner, L., & Juang, B. H. (1993). Fundamentals of Speech Recognition. Prentice-Hall.
- Decadt, B., et al. (2002). Phoneme-to-Grapheme conversion for out-of-vocabulary words. IEEE Workshop on Speech Synthesis.