SIGUL 2026 · 2026 · Workshop
Emmett Strickland et al. (INALCO / Paris-Saclay University) · → Paper · Demo: ? · Code: ?
A controlled pilot experiment that fine-tunes English, French, and Yoruba VITS base models on 30 minutes of Nigerian Pidgin (English-lexified) and Guadeloupean Creole (French-lexified) speech, finding that the French base model scores best for both creoles while the Yoruba substrate model is weakest, with statistical significance reached in only one of four comparisons.
Problem
Creole languages are uniformly low-resource for speech technology, and transfer learning from a better-resourced related language is a common way to build TTS with very little data. For creoles, “related” is ambiguous: the lexifier language (often European) supplies most of the vocabulary, while substrate languages (often African) are thought to shape syntax, phonology, and in some West African creoles lexical tone. Prior transfer work on creoles, such as Haitian ASR fine-tuned from French versus English, has compared lexifiers only and left substrate languages untested. The paper asks which parent language makes the better pretraining source for creole TTS.
Method
The authors recorded two new single-speaker datasets, each about 30 minutes of read prose from a female native speaker in a studio setting at 22,050 Hz: Nigerian Pidgin (Naijá, pcm, English lexifier) and Guadeloupean Creole (gcf, French lexifier). Recordings were segmented at sentence level, long sentences split into breath groups, and transcripts manually corrected. The data was collected for internal research use only (§2.1).
Three female single-speaker base models were trained with the CoquiTTS implementation of VITS: English on LJ Speech, French on the neutral-emphasis subset of SIWIS, and Yoruba on the TTS portion of IroyinSpeech (authors’ spelling ÌròyìnSpeech). Yoruba serves as a substrate for Naijá and as a typological proxy (isolating morphology, dense lexical tone) for the Fon-type substrates of Guadeloupean Creole. To control for data quantity, each base model used 3,000 randomly sampled utterances, 1,000 epochs, batch size 32, default CoquiTTS hyperparameters, and grapheme inputs; VITS was chosen because it needs no phoneme alignments (§2.2). Each base model was then fine-tuned for 600 epochs on 300 samples from each creole, giving six models (Table 1).
Evaluation used the remaining roughly 10% of each creole dataset as a reference set. The authors computed Mel Cepstral Distortion (MCD, via pymcd) and PESQ (downsampled to 16 kHz) between synthesized and reference utterances, compared models with Friedman tests, and ran an informal subjective evaluation consisting of their own listening impressions of intelligibility and naturalness (§2.3).
Key Results
All metric values below are means read from the box plots in Figures 1 to 4 (lower MCD and higher PESQ are better).
| Target creole | Metric | English base | French base | Yoruba base | Friedman p |
|---|---|---|---|---|---|
| Nigerian Pidgin | MCD | 17.36 | 16.98 | 17.33 | 0.22 |
| Nigerian Pidgin | PESQ | 1.06 | 1.08 | 1.05 | 0.51 |
| Guadeloupean Creole | MCD | 17.69 | 17.25 | 17.78 | 0.12 |
| Guadeloupean Creole | PESQ | 1.08 | 1.14 | 1.09 | 0.0006 |
The French base model has the best mean on all four comparisons, including for the English-lexified Nigerian Pidgin, contrary to the authors’ lexifier hypothesis. Only the Guadeloupean PESQ difference is statistically significant (§3.2, Figure 4). On Nigerian Pidgin, English and Yoruba are nearly tied on MCD, with English showing much wider spread (§3.1, Figure 1). On Guadeloupean Creole PESQ the Yoruba model edges the English one (1.09 versus 1.08 in Figure 4; the prose sentence in §3.2 lists the two numbers in reversed order but states the same ranking). French outputs also show a markedly higher upper quartile on PESQ for both creoles (Figures 2 and 4).
The authors’ listening impressions broadly agree: all Naijá models were intelligible but artificial; the English model handled consonant clusters well but buzzed; the Yoruba model had the closest vowel quality but degraded at consonant clusters (repeated syllables, epenthetic vowels, pauses); the French model was judged most natural. For Guadeloupean Creole the French model was clearly the most natural and the Yoruba model the weakest, and all Guadeloupean models were judged more natural than any Naijá model (§3.1, §3.2). These impressions are not quantified and come from the authors themselves.
Absolute quality is low across the board: PESQ means between 1.05 and 1.14 sit near the bottom of the scale, and MCD values around 17 are high, so the comparisons are between six weak systems.
Novelty Assessment
The contribution is the experimental question rather than any method: comparing lexifier against substrate pretraining for creole TTS under matched data budgets, which prior creole transfer work (lexifier-only, ASR) had not done. The modeling recipe is standard off-the-shelf VITS fine-tuning. The controls are careful for a pilot (equal base-model sample counts, same speaker gender, same recording equipment for both creoles), but the design has one base model per language, one speaker per creole, two reference-based objective metrics, and no formal listening test, so it cannot separate language effects from dataset or speaker effects. The authors are explicit that the results are partially inconclusive.
Field Significance
low — This is a small pilot that provides an early cautionary data point for low-resource transfer: in this setup, the historically related source language (lexifier or substrate) did not reliably give the best TTS fine-tuning starting point, and the authors point to phonotactic compatibility and orthographic depth as candidate factors. Its value is mainly as a hypothesis generator for creole and other low-resource TTS work.
Claims
- complicates: For low-resource TTS fine-tuning, the historically related source language of a creole (its lexifier or substrate) is not necessarily the best pretraining source.
Evidence: A French-pretrained VITS model gave the best mean MCD (16.98 versus 17.36 English, 17.33 Yoruba) and PESQ (1.08 versus 1.06, 1.05) for English-lexified Nigerian Pidgin, though neither difference was significant (Friedman p = 0.22 and 0.51). (§3.1, Figure 1, Figure 2)
- supports: Fine-tuning from a lexifier-language TTS model can benefit a creole whose orthography and vocabulary derive from that lexifier.
Evidence: For French-lexified Guadeloupean Creole, the French base model reached the highest mean PESQ (1.14 versus 1.08 English and 1.09 Yoruba), the only statistically significant difference in the study (Friedman p = 0.0006), and had the lowest mean MCD (17.25 versus 17.69, 17.78; p = 0.12). (§3.2, Figure 3, Figure 4)
- complicates: Phonotactic mismatch between a pretraining language and the target can outweigh segmental similarity in cross-lingual TTS transfer.
Evidence: Yoruba-based models, whose source language prefers V and CV syllables, repeatedly failed on consonant clusters in both creoles (repeated syllables, epenthetic vowels, unnatural pauses) even though the Yoruba-based Naijá model had the closest vowel quality, in the authors’ informal listening. (§3.1, §4)
- refines: With grapheme inputs and very small fine-tuning sets, target-language orthographic depth may affect achievable TTS quality as much as the choice of pretraining language.
Evidence: All Guadeloupean Creole models (shallow orthography) were judged more natural than any Nigerian Pidgin model (opaque English-derived spellings) with the same 300-sample fine-tuning budget, though the authors also note a speaker reading-experience confound. (§3.2, §4)
Limitations and Open Questions
Small, confounded pilot
Each language is represented by a single base model and each creole by a single speaker and roughly 30 minutes of audio; three of four model comparisons are not statistically significant, and the subjective evaluation is the authors’ own informal impressions with no reported scores. The French advantage could stem from acoustic properties of the SIWIS dataset rather than language, as the authors acknowledge.
- MCD and PESQ are reference-based signal metrics and all six systems score very low on PESQ; no intelligibility metric (WER/CER) or formal MOS test is reported.
- Base models themselves were not evaluated, so differences in base-model quality are not controlled.
- The two creole datasets differ in speaker reading experience as well as orthography, confounding the cross-creole comparison.
- Grapheme inputs entangle phonology with orthography; the authors propose repeating the experiment with phonetized text, and note that Yoruba spelling encodes tone while Naijá spelling does not.
- The recorded creole datasets are for internal research use and are not released; code and samples are not reported.
Wiki Connections
- GAN Vocoder — All six systems are VITS models whose adversarially trained waveform decoder is pretrained and then fine-tuned end to end on 300 creole utterances.