📄 Transfer Learning for Creole TTS: A Pilot Study on Whether Substrate Phonologies or Lexifier Vocabularies Matter More

Summary

Creole languages typically take their vocabulary from a lexifier and much of their phonology and syntax from substrate languages. This raises a practical question: which parent language gives a better starting point for transfer learning in text-to-speech? This pilot study tests that question under matched low-resource conditions for two creoles: Nigerian Pidgin (Naijá, pcm), which is English-lexified, and Guadeloupean Creole (gcf), which is French-lexified.

The authors recorded about 30 minutes of studio read speech from one female native speaker per language. They then fine-tuned VITS base models, built with CoquiTTS on grapheme input and trained on English, French and Yoruba, which produced six systems. Evaluation combined MCD, PESQ and informal listening.

French-based models came out ahead for both creoles, but the difference was statistically significant only for Guadeloupean Creole. Yoruba-based models were generally the weakest, mainly because they handled consonant clusters poorly. The authors suggest that historical relatedness may predict transfer less well than phonotactic compatibility and orthographic depth. This is encouraging for languages that lack a close, well-resourced relative.

Key points

• New datasets: two single-speaker creole TTS datasets (~30 min each, 22,050 Hz, RØDE Wireless GO II microphone).
  • Recordings were segmented into sentences or breath groups.
  • About 10% of each dataset was held out as a gold-standard test set.
  • Speaker consent covers internal research use only.
• Base models: trained on LJ Speech (English), SIWIS (French, "neutral" emphasis subset only) and ÌròyìnSpeech (Yoruba).
  • Base training used 3000 samples, 1000 epochs and batch size 32.
  • Fine-tuning used 300 creole samples for 600 epochs.
• Naijá results: French gave the best mean MCD (English 17.36, Yoruba 17.33) and the highest mean PESQ (French 1.08, English 1.06, Yoruba 1.05).
  • Friedman tests were not significant (p = 0.22 for MCD, p = 0.51 for PESQ).
• Guadeloupean Creole results: a Friedman test gave p = 0.0006, a significant difference between models.
  • The gcf models outperformed their pcm counterparts overall.
  • The authors attribute this partly to gcf's shallow orthography and to the gcf speaker's experience reading creole texts.
• Future work: the authors plan to repeat the experiment with phonetized input to separate the effects of phonology and orthography. They also want to study how tone marking in spelling interacts with prosody and to evaluate the base models themselves.