Natural language processing (NLP) and speech technologies have made significant progress in recent years; however, they remain largely focused on standardized language varieties. Dialects, despite their cultural significance and widespread use, are underrepresented in linguistic resources and computational models, resulting in performance disparities. To address this gap, we introduce Saar-Voice, a six-hour speech corpus for the Saarbrücken dialect of German. The dataset was created by first collecting text through digitized books and locally sourced materials. A subset of this text was recorded by nine speakers, and we conducted analyses on both the textual and speech components to assess the dataset's characteristics and quality. We discuss methodological challenges related to orthographic and speaker variation, and explore grapheme-to-phoneme (G2P) conversion. The resulting corpus provides aligned textual and audio representations. This serves as a foundation for future research on dialect-aware text-to-speech (TTS), particularly in low-resource scenarios, including zero-shot and few-shot model adaptation.
Figures & tables
Figure 1: The isogloss dat-das-line distinguishing areas in which Rhine Franconian dialects (“das”) and Moselle Franconian dialects (“dat”) are spoken ( Drenda, 2008 ) in Saarland (smaller area in the south-west marked with thin border lines) and Rhineland-Palatinate.
Source
Domain
# Sent
# Words
# U. Words
Printed Books
An da Saar gefonn
Poetry
698
6,228
1,978
Geschaffd - Gelääbd
Poetry; Prose
394
3,302
1,214
Saa, was de willschd
Prose; Poetry
2,216
19,711
3,147
Was wääs dann isch
Poetry; Prose
2,525
18,445
3,490
Locally Sourced
Table 1: Textual corpus statistics of the collected texts for the Saarbrücken dialect.
Speakers
Total speakers
9
Gender ratio
4F / 5M
Speech
Recorded sentences
4,871
Total duration (hh:mm:ss)
05:54:48
Avg. duration per sent. (sec)
4.37
Table 2: Saar-Voice Corpus Statistics.
Speaker
Gender
Total audio (hh:mm:ss)
# sentences
Mean F0
Median F0
F0 range (5–95%)
W/s
P01
F
00:23:17
301
203
200
141-278
1.8249
P02
F
01:02:28
1075
251
249
203-312
2.4166
P03
M
01:15:37
899
115
107
80-176
1.7352
P04
M
00:27:56
400
143
140
100-196
2.2056
P05
F
00:27:18
399
213
212
175-256
2.0203
P06
M
00:18:43
300
119
115
84-160
2.3334
Table 3: Speaker-level acoustic statistics. F0 in Hz. W/s = words per second.
Figure 2: SNR distribution on Saar-Voice.
Word
English
Epitran
Ours
berechne
calculate (Imp.)
\textipa b \textschwa \textscr \textepsilon xne \textlengthmark
\textipa b \textschwa \textscr \textepsilon xn \textschwa
bidde
please
\textipa b \textsci tde \textlengthmark
\textipa b \textsci d \textschwa
Midde
middle
\textipa m \textsci tde \textlengthmark
\textipa m \textsci d \textschwa
odder
or
\textipa \textopeno td \textschwa \textscr
\textipa \textopeno d \textturna
buuch
book (Imp.)
\textipa bu \textlengthmark u \textlengthmark x
\textipa bu \textlengthmark x
ääner
one
\textipa æ \textlengthmark æ \textlengthmark n \textschwa \textscr
\textipa æ \textlengthmark n \textturna
Table 4: Examples of G2P Errors. Corrections of transcriptions are estimates by a native speaker.
School of Artificial Intelligence, University of Chinese Academy of Sciences · Institute of Automation, Chinese Academy of Sciences · Wuhan AI Research +1