LREC 2026 · 2026 · Conference
Adina Ioana Vladu et al. (Universidade de Santiago de Compostela) · → Paper · Demo: ? · Code: ?
Releases an 18-hour, CC BY 4.0, professionally recorded single male speaker Galician TTS corpus whose text was redesigned for better sentence-modality balance and diphone coverage than its female predecessor, and validates it by training VITS and Matcha-TTS with a 34-listener MOS test.
Problem
Galician is a lesser-resourced language with reasonable text resources but few open, high-quality speech corpora for neural TTS. The earlier open resources are the Cotovía system, the multi-speaker CRPIH_UVigo-GL-Voices dataset (about 21 hours), and the same project’s female single-speaker corpus Nos_Celtia-GL. The authors identify three weaknesses in Nos_Celtia-GL that they set out to fix: an imbalanced distribution of sentence modalities (few exclamatory and imperative sentences), lexical repetition in numeric and formulaic phrases, and under-representation of certain phonetic contexts. A male voice is also needed to complement the existing female one.
Method
The contribution is the corpus and its design procedure. Speaker selection used a perceptual test with 37 linguists and speech-technology specialists, rating three candidate speakers on clarity, pronunciation, pleasantness, intonation and overall quality; the selected professional voice talent scored highest on pronunciation, intonation and the combined score (Figure 1). The authors also describe the speaker’s acoustic profile: a speaking rate of about 2.8 words per second, a mean F0 of 172 Hz, and consistent mid-open/mid-close vowel contrasts.
The text corpus has 16,121 sentences (168,320 words) drawn from parliamentary transcripts, public-broadcaster scripts, dictionary examples, an oral-language archive, a prior GTM/CRPIH TTS corpus, literary works chosen for specific phonetic and prosodic contexts, and a 500-word subcorpus selected with CorpusCrt from a pronunciation dictionary to guarantee diphone coverage (Table 1). All text was normalised to current Galician orthographic norms. Sentence lengths approximate a Gaussian centred around 10 to 11 words, with a deliberate surplus of one-word sentences. The corpus was intentionally made smaller than Nos_Celtia-GL because earlier work by the same group found that about 13,000 sentences were enough for high-quality results.
Recordings were made in a professional studio (Neumann U-87, 48 kHz) under real-time supervision by linguists, with retakes where needed. The release has three versions: raw, edited (mouth noises, clicks and breaths removed, with 600 ms of leading and trailing silence), and edited plus normalised to -10 dBTP for direct use in training.
To validate the corpus, the authors train two existing TTS systems from scratch on grapheme inputs, using the Cotovía front end for text normalisation. The first is VITS, an end-to-end conditional VAE with normalising flows and an adversarially trained HiFi-GAN-style decoder, trained for 1,300 epochs. The second is Matcha-TTS, a non-autoregressive encoder-decoder trained with optimal-transport conditional flow matching, paired with a HiFi-GAN vocoder, trained for 4,000 epochs with the authors’ default settings. The paper does not say whether the Matcha-TTS vocoder was fine-tuned on the new corpus or used pretrained. No model sizes are reported.
Key Results
The corpus is well balanced by its own statistics. Compared with Nos_Celtia-GL, interrogatives rise from 12.06% to 15.02%, exclamatives from 4.3% to 11.51%, and declaratives fall from 80.28% to 67.07% (Table 2). The corpus contains 38 phoneme types and 947 diphone types, 99.75% of the diphones occur at least 30 times, and 98% of phonemes occur more than 5,000 times (§2.3.2, Table 3).
The subjective test had 34 listeners (29 native Galician speakers and five high-proficiency speakers), each rating 15 stimuli (five randomly sampled per condition) on separate five-point quality and naturalness scales. The 21 unseen test sentences were stratified by sentence type and included one hard spontaneous-speech utterance (Figure 2):
| System | Quality MOS | Naturalness MOS | UTMOSv2 |
|---|---|---|---|
| Original recordings | 4.65 | 4.37 | 3.66 |
| VITS | 3.99 | 3.74 | 3.22 |
| Matcha-TTS | 3.89 | 3.38 | 3.55 |
VITS is significantly better than Matcha-TTS on naturalness. The quality difference is not significant at the 95% level. The authors compare against the best model from their earlier Nos_Celtia-GL study (3.74 naturalness, 4.19 quality), where natural speech scored 4.10 and 4.37. Against that, VITS on Nos_Brais-GL matches on naturalness and is 0.20 lower on quality. Because the male recordings score higher, the gap between synthetic and natural speech is larger here: 0.66 for quality and 0.63 for naturalness, against 0.18 and 0.36 for the female corpus (§4.1). These comparisons come from separate listening tests with different listener pools, so the “statistically identical” phrasing applies to absolute scores across tests, not to a paired comparison.
The objective ranking from UTMOSv2 does not match the human ranking. It puts Matcha-TTS (3.55) above VITS (3.22), and it rates the natural recordings far lower than listeners did (3.66 against 4.37 subjective naturalness) (§4.2, Figure 3). The authors attribute this to UTMOSv2 being trained mostly on English data (BVCC, SOMOS). They do not test that explanation directly.
Novelty Assessment
This is a resource paper. The scientific content lies in the corpus-design methodology, which adapts the group’s earlier Nos_Celtia-GL procedure to fix specific, measured shortcomings (modality balance, diphone coverage, lexical repetition), and in documenting that procedure transparently. The TTS experiments only check that the data is usable. Both models are trained with stock or lightly specified hyperparameters, and the authors say both probably need corpus-specific tuning. Nothing is new in the modelling. The UTMOSv2 mismatch is a useful small cautionary observation but rests on three systems and a small test set.
Field Significance
Moderate. The paper adds an openly licensed, professionally recorded male voice corpus for a lesser-resourced language, with a documented and reproducible design procedure that serves as a worked template for phonetically and modally balanced single-speaker TTS corpora. Its evaluation also gives a concrete data point that an English-trained MOS predictor can invert the human ranking of systems in another language.
Claims
- supports: Single-speaker corpora of under 20 hours with deliberately balanced phonetic and sentence-type coverage are sufficient to train end-to-end neural TTS that listeners rate well above the midpoint for naturalness in a lesser-resourced language.
Evidence: VITS trained from scratch on the 18-hour Nos_Brais-GL corpus reaches 3.99 quality and 3.74 naturalness MOS from 34 listeners, on a test set that includes interrogative, exclamatory, elliptical and spontaneous-speech sentences. (§3, §4.1, Figure 2)
- complicates: Automatic MOS predictors trained mostly on English synthetic speech can be miscalibrated for other languages, to the point of reversing the human ranking of TTS systems.
Evidence: UTMOSv2 scores Matcha-TTS (3.55) above VITS (3.22), while listeners significantly prefer VITS on naturalness (3.74 against 3.38), and UTMOSv2 rates natural Galician recordings at 3.66 against 4.37 subjective naturalness. (§4.2, Figure 3, Figure 2)
- complicates: The gap between synthetic and natural speech in MOS depends on how listeners rate the source voice, so similar absolute synthetic scores can hide very different shortfalls from the original recordings.
Evidence: VITS on the male corpus matches the earlier female-corpus best model on naturalness (3.74), but the higher ground-truth ratings for the male recordings (4.65 and 4.37, against 4.37 and 4.10) widen the gap to 0.66 and 0.63, against 0.18 and 0.36 for the female corpus. (§4.1)
- refines: When the training corpus is the only thing that changes, a jointly trained VAE-GAN end-to-end model can be rated more natural than a flow-matching acoustic model with a separate vocoder, even when the two are rated about equal on quality.
Evidence: VITS beats Matcha-TTS significantly on naturalness (3.74 against 3.38), but the quality difference (3.99 against 3.89) is not significant at 95% confidence. Matcha-TTS used default hyperparameters, which the authors say may need tuning for this corpus. (§3, §4.1, Figure 2)
Limitations and Open Questions
- The corpus has a single male speaker reading in a neutral style with moderate expressiveness. It supports single-voice TTS but not multi-speaker, zero-shot or expressive synthesis.
- The validation listening test is small: 21 test sentences, with each listener hearing only five randomly sampled stimuli per condition. Only one exclamatory and one elliptical sentence are included, so the test says little about whether the improved modality balance actually helps synthesis of those sentence types.
- The paper does not compare models trained on Nos_Brais-GL with models trained on Nos_Celtia-GL in the same listening test. The cross-corpus comparisons rest on separate tests with different listeners and different voices, so they cannot isolate the effect of the corpus-design changes.
- Neither model was tuned for the corpus, and the paper does not say whether the Matcha-TTS HiFi-GAN vocoder was adapted to the new speaker, which may account for part of Matcha-TTS’s lower naturalness.
- No objective intelligibility metric (WER or CER) is reported. The proposed explanation for the UTMOSv2 miscalibration (cross-lingual mismatch) is plausible but untested.
Wiki Connections
- GAN Vocoder — VITS is trained from scratch on the corpus with its adversarially trained HiFi-GAN-style waveform decoder, and Matcha-TTS uses a HiFi-GAN vocoder to produce waveforms.
- Flow Matching — Matcha-TTS, an OT-CFM acoustic model, is one of the two validation baselines and is rated less natural than VITS on this corpus.
- Subjective Evaluation — a 34-listener online MOS test with separate quality and naturalness scales provides the main evidence that the corpus is usable.
- Evaluation Metrics — UTMOSv2 reverses the human ranking of VITS and Matcha-TTS on Galician speech and badly underrates natural recordings, a cross-lingual calibration warning for automatic MOS predictors.
- HiFi-GAN — the vocoder design used inside VITS and as the waveform generator for Matcha-TTS in the validation experiments.