1 of 15

Saar-Voice: �A Multi-Speaker Saarbrücken Dialect Speech Corpus

Dialects in NLP – A Resource Perspective LREC 2026, Palma, Mallorca (Spain)

Lena Oberkircher – s8leober@stud.uni-saarland.de

Jesujoba O. Alabi

Dietrich Klakow

Jürgen Trouvain

2 of 15

2

[01] https://commons.wikimedia.org/wiki/File:Europe_blank_laea_location_map.svg

[01]

3 of 15

3

[02] https://als.wikipedia.org/wiki/Datei:Germany_location_map.svg

[02]

4 of 15

Dialectal regions

Moselle Franconian
Rhine Franconian

4

[03] https://www.regionalgeschichte.net/rheinhessen/sprache/dialektatlas-rlp-saar.html

Moselle Franconian

Rhine Franconian

[03]

5 of 15

Why this dialect?

Spoken by ~ the entire population, but has virtually no presence in media, research or language technologies
UNESCO classifies Rhine and Moselle Franconian as vulnerable
Research can help bring awareness and prestige
No existing speech corpus to use for research -> we built one

5

6 of 15

Text Dataset

6

Source	Domain(s)	# Sentences	# Words
Printed Books	Poetry, Prose	5833	47686
Locally Sourced	Prose, Folktale, Poetry	2838	26789
Localized Translations	Localized Data	101	805
Total		8772	75280

7 of 15

Localized Translations

7

German Original	Bitte plane ein Treffen mit Petra in Wiesbaden am Mittag.
Entities Replaced	Bitte plane ein Treffen mit Anna in Oberbexbach am Mittag.
Translated + Localized	Bidde plaan e Dreffe midd Anna in Owwerbeddschbach am Middaach.

MASSIVE Dataset (German split) [04]

[04] https://aclanthology.org/2023.acl-long.235/

8 of 15

Audio recordings

8

Number of speakers	9
Gender ratio	4F/5M
Recorded sentences	4871
Total duration (hh:mm:ss)	05:54:48

9 of 15

Audio quality

SNR: Majority of samples fall in clean audio threshold

9

10 of 15

Ongoing work

Adaptation of multilingual TTS models (all exposed to Standard German):

ZMM-TTS - github.com/nii-yamagishilab/ZMM-TTS#
XTTS - huggingface.co/coqui/XTTS-v2
YourTTS - huggingface.co/AOLCDROM/YourTTS-Fr-En-De-Es

Unseen speaker / dialect experiments (“zero-shot inference”)

10

11 of 15

Unseen speaker / dialect inference

→ promising but inconsistent results across models

11

model	MCD (mean)	Speaker sim. (mean)	NISQA (mean)
XTTS	13.5285	0.7737	3.6726
YourTTS	6.9804	0.7676	3.6591
ZMM-TTS	8.4939	0.5682	3.1133

12 of 15

Future work

Our part:

Fine-tuning of the models to improve performance significantly
Evaluate results with human listening test

Hypothetical work

Extension of the dataset
Other speech technologies: ASR, voice assistants, chatbots, …

12

13 of 15

Thank you!

Questions? Comments?

-> s8leober@stud.uni-saarland.de

13

CREDITS: This presentation template was created by Slidesgo, and includes icons, infographics & images by Freepik

??

Why not train your own model?

Focus is adaptation; training from scratch is not feasible

Is SL truly unseen?

Yes at dialect level, as none of the models have seen SL speech or text
But: different from truly unseen languages: will probably benefit from Standard German
Dialect is distinct enough from Standard German in my opinion

How is unseen dialect TTS different from unseen language TTS?

For unseen languages, models often lack phonetic coverage
For dialects, the model (often) has a prior knowledge of the Standard Language and may treat dialects as noise
The challenges are different: for dialects, we may (unknown) see the model normalize to the Standard Language

Why focus on multilingual models over dialect-specific models?

Multilingual models offer more scalable alternative
Because dialect-specific models often require substantial linguistic resources which are not widely available

Why (the) three models?

Represent different strategies:
ZMM-TTS allows for voice-cloning-zero-shot; MMS focuses more on explicit zero-shot; XTTS focuses on massive multilingual pretraining
Explores which kind of model is best suited for dialect adaptation

What if fine-tuning doesn’t help?

Shows the limits of current models
Shifts the focus to data collection and finding new strategies

What if fine-tuning hurts voice cloning capabilities?

Important finding that highlights the trade-off between dialect adaptation and speaker generalization
One goal of thesis: discovering and understanding such trade-offs

How is MMS different?

No inference-time speaker conditioning
Uses language-specific models with fixed speaker identity

Selective Fine-tuning?

Refers to adapting only the components of a pre-trained model that control a specific part of the task (e.g. for TTS, only pronunciation and prosody) while keeping the rest frozen
Suitable for low-resource dialect adaptation, as we can keep low-level waveform generation, speaker representations etc. fixed
Not enough data for full training; might even be too little for full fine-tuning (will have to explore)

Is dialectal correctness subjective?

Yes, and we will conduct a listening test with multiple participants (as many as we can find with our limited resources)

14 of 15

Icon pack

14

15 of 15

Alternative resources

Here’s an assortment of alternative resources whose style fits that of this template:

Flat design geometric background

15