Listen to each model
One sentence of the fable "The North Wind and the Sun", a text no model saw in training. Scores are the mean opinion of 23 listeners, out of 5, with 95% intervals.
Fine-tuned from French, graphemesBest mean for both voices, 13 s
4.10
Fine-tuned from French, phonemesPhoneme input, 12 s
3.96
Walloon data only, graphemesTrained from scratch, 13 s
3.38
Walloon data only, phonemesTrained from scratch, 11 s
2.70
Natural speechReference recording, not published
4.55
Text, grapheme models
Li bijhe et l’ solea estént ki s’ margayént po sawè kî çki, des deus, esteut l’ pus foirt. Mins ç’ côp la, la k’ i veyèt on tchminåd k' arivéve pyim piam, dins on bea noû tchôd paltot.
Input, phoneme models
li biːç ɛ l sɔlja ɛstẽ ki s maʁɡajẽ pɔ sawɛ kiː ski , dɛ døː , ɛstøː l py fwaʁ . m ɛ̃ s koː la , la k i vɛjɛ õ tsminɔː k aʁivef pjim pjam , d ɛ̃ õ bja nuː tsoː paltɔ .
Starting point
The French model before fine-tuning, 7 s
La bise et le soleil se disputaient, chacun assurant qu'il était le plus fort, quand ils ont vu un voyageur qui s'avançait, enveloppé dans son manteau.
The problem
Walloon is a Gallo-Romance language spoken in Belgium and the French Ardennes, with no speech synthesis system. The data: The Little Prince read in standardised Walloon, 156 minutes by a male speaker and 18 minutes by a female speaker.
What I did
- Prepared the corpus: segmentation, resampling, normalisation, pause reduction.
- Trained VITS from scratch and fine-tuned a French model.
- Compared grapheme input with phonemes from the group's rule-based converter.
- Ran objective metrics and a 23-listener perceptual test, then wrote the paper.
Results
Fine-tuning from French gave the best mean for both voices, 4.10/5, against 4.48 and 4.55 for natural speech. With 18 minutes of data it lifted the female voice from 3.38 to 4.10; with 156 minutes the male models were not significantly different. The automatic MOS predictor (UTMOSv2) did not transfer to Walloon.