NoDaLiDa/Baltic-HLT 2025 · 2025 · Workshop
Kiissel et al. (Institute of the Estonian Language) · → Paper · Demo: ✓ · Code: ✓
A dedicated TTS system for Estonian isolated-word synthesis that achieves correct homograph disambiguation for palatalisation and quantity contrasts by requiring users to supply explicit diacritics, bypassing the context-dependent morphological disambiguation that makes standard Estonian TTS systems unreliable on single words.
Problem
Standard Estonian TTS pipelines rely on statistical morphological disambiguation trained on sentence context, which makes them unable to resolve homographs when input is a single word. Estonian has a particularly large homograph space: over 756 orthographic forms ambiguous in palatalisation and more than 22,000 ambiguous in quantity (second vs. third quantity degree). For learners’ dictionaries, pronunciation training, and speech therapy applications, where words are presented in isolation, a 50% chance of producing the wrong pronunciation variant is unacceptable. No prior Estonian TTS system addressed this constraint directly.
Method
The system is built on the Merlin TTS toolkit, a deep neural network statistical parametric synthesis framework developed at CSTR Edinburgh. Training data consists of 31,215 professionally recorded isolated Estonian words (approximately 10 hours 36 minutes) drawn from the Estonian Combined Dictionary (CombiDic), covering broad phonetic contexts and recorded in studio conditions. Because the training utterances are isolated words, the model learns prosody appropriate to word-level citation form rather than sentence-embedded speech.
Pre-processing skips the standard morphological parsing stage entirely. Instead, users supply the orthographic input with explicit diacritics that encode the phonological features normally assigned by the Vabamorf parser: third quantity (Q3), palatalisation, irregular lexical stress, and compound boundaries. These diacritics follow the annotation conventions used in Estonian dictionaries and the Vabamorf system, so the mapping from user-supplied annotation to phoneme string is deterministic. Grapheme-to-phoneme conversion then proceeds using the standard Estonian TTS pipeline. The system is deployed as a public web interface where users enter annotated words and download WAV files of the synthesised pronunciation.
Key Results
A perception test with eight phonetics and linguistics experts evaluated 32 homograph pairs distinguishing palatalisation and 32 pairs distinguishing quantity. For palatalisation, 26 of 32 items were identified correctly by all eight evaluators (100% agreement); the remaining six items achieved 75-88% correct recognition, with failures concentrated on palatalised /t/ and /tj/ contrasts. This reflects a known variability in Estonian palatalisation perception rather than a synthesis deficiency. For quantity, 31 of 32 items were unanimously recognised correctly; only one Q2 form failed for a single evaluator. No baseline comparison against a natural speech recording is reported, nor is a MOS naturalness evaluation. The paper contrasts the system’s near-perfect disambiguation rate against the 50% chance-level performance of standard Estonian TTS systems on homographic inputs.
Novelty Assessment
The contribution is primarily engineering integration applied to a narrow but practically important niche. Merlin is a well-established statistical parametric synthesis toolkit (2016), and the pre-processing strategy of shifting disambiguation responsibility to the user rather than the model is straightforward. The genuine contribution is the assembly of a purpose-built training corpus of isolated-word recordings and the decision to treat user-supplied diacritics as the disambiguation mechanism rather than attempting automatic context-free homograph resolution. This is an honest and functional solution to the problem, but it transfers linguistic knowledge requirements to the user rather than solving disambiguation algorithmically. The evaluation covers only expert evaluators (n=8) on a controlled homograph subset; no naturalness MOS or intelligibility test is reported, and no comparison to natural speech is provided.
Field Significance
Low — This paper addresses a real gap in Estonian language technology for pedagogical and speech therapy applications, but the approach is a targeted application of existing tools to a resource-constrained low-resource language setting. The system provides useful functionality, but the technical contribution does not generalise beyond the specific homograph problem in Estonian. It is primarily confirmatory evidence that statistical parametric synthesis (Merlin) remains viable for low-resource isolated-word scenarios when combined with an annotated training corpus.
Claims
-
supports: Explicit phonological annotation in TTS input can substitute for context-dependent disambiguation when morphological parsing is unavailable or unreliable.
Evidence: The system achieves near-perfect expert recognition of palatalisation and quantity contrasts in Estonian homographs by requiring users to supply diacritics rather than relying on statistical disambiguation, compared to 50% chance-level performance in standard Estonian TTS systems. (§3.2)
-
complicates: Perception-based accuracy metrics alone are insufficient for evaluating isolated-word TTS quality.
Evidence: The evaluation uses expert phoneticians (n=8) on a controlled set of 64 homograph pairs with no naturalness MOS, no intelligibility test with naive listeners, and no comparison to natural speech recordings; the scope of claims about “very good” performance is therefore limited to pronunciation accuracy on a specialist-curated test. (§3.1, §3.2)
-
supports: Purpose-built isolated-word training corpora enable statistical parametric TTS systems to generate citation-form prosody that is inappropriate when systems are trained on continuous speech.
Evidence: Training on 31,215 professionally recorded isolated Estonian words (10h 36min) from CombiDic yields utterance prosody judged suitable for pronunciation training and dictionary applications. (§2.1, §4)
Limitations and Open Questions
The user interface requires manual diacritic insertion, which creates a significant usability barrier for non-linguists. The paper acknowledges this and plans automatic tagging with multiple homograph options as future work. More substantively, the evaluation relies entirely on expert listeners; it is unknown whether the synthesised contrasts are perceptible to the target users (language learners, speech therapy clients) who may have less phonological sensitivity. There is no naturalness evaluation alongside the accuracy test, so the system’s overall speech quality relative to other Estonian TTS systems or to natural speech is not established. The choice of Merlin (a 2016-era toolkit) is noted as intentional for data efficiency and control, with more advanced TTS techniques planned for future systems.
Wiki Connections
- Evaluation Metrics — the paper uses a perception test with expert evaluators as the primary evaluation method, foregoing standard MOS naturalness metrics, which limits the scope of quality claims.
- Prosody Control — the system achieves utterance-level prosody naturalness for isolated words by training exclusively on citation-form recordings, rather than through an explicit prosody control mechanism.