When Spelling Leaves Tone Out: What Morpho-VITS Adds to Kinyarwanda TTS

TL;DR for operators

Morpho-VITS¹ improves Kinyarwanda text-to-speech on both automatic and native-listener evaluation. Compared with a VITS baseline, word error rate falls from 13.4% to 9.4%, while perceived intelligibility rises from 92.1% to 96.7%. Listener scores also improve for naturalness and intonation.

The architectural change is explicit linguistic structure. Kinyarwanda spelling does not mark all of the tonal information needed for natural speech, so Morpho-VITS separately represents the meaningful grammatical and lexical pieces inside words and feeds that information into phoneme-level processing.

The deployment boundary is equally concrete. On sentences containing morphemes absent from training, Morpho-VITS reaches 19.3% WER, versus 8.3% when every morpheme was previously observed. Its 268 million parameters also substantially exceed the baseline’s 97 million, and training takes 149 hours rather than 96 hours in the reported four-GPU setup. The paper therefore supports morphology-aware modeling as a promising design choice in this Kinyarwanda setting, but it does not isolate morphology from additional model capacity.

A voice can pronounce the words and still realize them incorrectly

Text-to-speech systems do not receive every property of speech directly from written text. A written sentence may identify the words clearly while leaving out information that determines how those words should sound.

Kinyarwanda makes this problem particularly visible. Pitch patterns can affect correct linguistic realization, yet standard orthography does not explicitly mark all of those tonal cues. Their realization depends partly on lexical, morphological, and syntactic structure. A system that works only from phoneme-like units therefore has to recover information that is not fully present in its immediate input.

The Morpho-VITS results show that changing the text representation can matter. On 10,000 Kinyarwanda news sentences, the proposed system reduces character error rate from 3.0% to 1.7% and word error rate from 13.4% to 9.4%. In native-listener evaluation, naturalness rises from 3.33 to 3.55, intonation from 2.95 to 3.12, and perceived intelligibility from 92.1% to 96.7%.

Those gains shift attention upstream. Better synthesis is not only a question of generating better acoustics from a fixed text representation. The representation itself can omit information the model needs.

Morpho-VITS gives phonemes access to meaningful word structure

A morpheme is the smallest meaningful grammatical or lexical unit inside a word. In a morphologically complex language, one written word can contain several such units, each contributing information about lexical meaning or grammatical structure.

Morpho-VITS retains the broader VITS/VITS2 speech-synthesis framework but changes how text is encoded. The system represents the phoneme-like sequence and the morpheme sequence separately. An eight-layer morpheme encoder produces representations of the morphological structure, while an eight-layer phoneme cross-encoder combines phoneme self-attention with phoneme-to-morpheme cross-attention.

That cross-attention is deliberately constrained. A phoneme can directly attend only to morphemes belonging to the same word. The attention score also includes a relative positional bias, giving the model information about where a morpheme sits relative to the phoneme being processed.

The design matters because it narrows what “adding linguistic structure” means. Morpho-VITS does not simply append a global morphological embedding to the sentence. It creates a controlled path through which phoneme representations can use the meaningful components of their own word before those representations condition the speech model.

The listener results support more than an ASR improvement

Automatic speech-recognition error is useful here because it provides a scalable intelligibility measure across 10,000 synthesized sentences. But lower ASR error alone would not establish that the resulting speech sounds more natural to people.

The paper therefore adds a paired human evaluation with 21 Kinyarwanda native speakers, producing 2,100 clip-pair ratings. The morphology-aware model scores higher not only on perceived intelligibility but also on naturalness and intonation.

The agreement between automatic and human evaluation strengthens the within-setting comparison. The model is not merely producing audio that happens to be easier for one ASR system to transcribe. Native listeners also report improvements in the speech dimensions that motivated the architecture.

This remains comparative evidence rather than a clean causal decomposition of the architecture. That distinction becomes important when deciding what to build next.

Unseen morphemes turn data coverage into a model requirement

About 9% of the automatic evaluation text contains morphemes absent from training. This subgroup functions as a generalization test rather than a second headline benchmark.

The degradation is substantial. Morpho-VITS records 2.9% CER and 19.3% WER when unseen morphemes are present, compared with 1.5% CER and 8.3% WER when all morphemes have been observed during training.

Human judgments move in the same direction. With unseen morphemes, naturalness falls to 3.44 from 3.56, intonation to 3.00 from 3.13, and perceived intelligibility to 93.8% from 97.0%.

For production systems, this changes the data question. If explicit morphology becomes part of the synthesis pipeline, corpus design cannot be judged only by total speech hours. Coverage of the morpheme inventory also becomes relevant. A system may have substantial acoustic training data while remaining exposed to failures in linguistically sparse regions of the text distribution.

The experiment does not isolate morphology from added capacity

The main interpretive constraint is model size. Morpho-VITS has 268 million parameters, compared with 97 million for the baseline. The morpheme encoder alone accounts for 169 million parameters.

The reported models use the same speech datasets, the same one-million-update schedule, batch size 32, and peak learning rate of $10^{-4}$. That makes the performance comparison informative within the reported setup. It does not, however, answer what would happen if the baseline received comparable additional capacity without morphological structure.

No capacity-matched ablation is reported. Nor is there an alternative morphology-integration ablation that would separate the value of the specific cross-attention design from the broader increase in representational capacity.

Compute requirements move in the same direction. On four RTX 5090 GPUs, reported training time is 96 hours for the baseline and 149 hours for Morpho-VITS.

The appropriate conclusion is therefore narrower than “morphology caused the gain.” The experiment shows that this morphology-aware architecture performs better than the reported VITS baseline. It does not determine how much of that improvement comes specifically from morphology rather than from the additional parameters and architectural changes used to represent it.

Speech product decisions should include linguistic coverage, not only model scale

Paper evidence. In the reported Kinyarwanda setting, the morphology-aware system improves automatic intelligibility and native-listener judgments, while performance deteriorates when the text contains unseen morphemes.

Cognaptus inference. For voice interfaces, IVR systems, and accessibility products in languages where ordinary writing omits pronunciation-critical information, explicit linguistic structure deserves evaluation as a first-class input design choice. Data collection may also need a coverage objective: not simply more recorded speech, but speech and text that expose the model to a broad range of productive morphological forms.

Uncertainty. The evidence comes from one language. It does not establish the same benefit for other tonal Bantu languages, languages with different morphological systems, or languages whose orthography already represents the relevant pronunciation cues. The model-capacity difference also prevents a clean estimate of the marginal value of morphology itself.

That leaves a practical research agenda for deployment: compare morphology-aware and capacity-matched alternatives, measure performance by morphological coverage, and determine whether the additional training and model cost buys enough improvement for the target speech application.

The representation can be part of the speech problem

Morpho-VITS makes a specific point about text-to-speech system design. When written input omits information needed for correct realization, increasing the sophistication of downstream speech generation may not be the only response. The text representation can be enriched with linguistic structure that exposes information the acoustic model would otherwise have to infer indirectly.

In Kinyarwanda, that design is associated with better intelligibility, intonation, and naturalness. It also creates a measurable dependency on morpheme coverage and arrives in a model almost three times the size of the baseline.

The paper therefore shifts the engineering question from whether morphology can be encoded to when its added structure, data requirements, and compute costs are justified by the speech quality it produces.

Cognaptus: Automate the Present, Incubate the Future.