arXiv · 2026 · Preprint

Lena S. Oberkircher et al. (Saarland University) · → Paper · Demo: ? · Code: ?

Releases Saar-Voice, a six-hour, nine-speaker read-speech corpus of the Rhine Franconian dialect spoken around Saarbrücken, built from digitized dialect books, community texts, and localized MASSIVE translations, with a text-side analysis of orthographic variation and Standard-German G2P failure modes but no TTS experiments.

Problem

German dialects are poorly represented in speech resources: existing German TTS corpora (MLS, CML-TTS, FLEURS-R, Thorsten-Voice) cover Standard German, and dialect efforts have targeted Swiss German, Viennese, Austrian varieties, Luxembourgish, and Southeast German groups. According to the authors, the Saarbrücken dialect has only a small Bible-derived subset in MMS. Dialects are not just accented Standard German: they differ in orthography-to-phonology mapping, morpho-syntax (article system, adjectival inflection), and lexicon (e.g. French loans), and they have no standardized spelling, so the same word is written several ways even by one writer (§1.2, §1.3). The paper sets out to build a high-quality multi-speaker corpus that could support dialect-aware TTS in low-resource, zero-shot, and few-shot adaptation settings (§2, §6). The arXiv comment states the paper was accepted at the DialRes workshop at LREC 2026.

Method

The contribution is a corpus-construction pipeline rather than a model.

Text collection (§3.1, Table 1). 8,772 sentences (75,280 whitespace tokens, 11,303 unique word forms) come from three sources: four printed dialect books (poetry and prose) digitized with an online OCR service (ocr.ac), chosen because other tools, including tesseract, misread the dialect diacritics ‘á’ and ‘ò’ as ‘ä’ and ‘ö’; OCR errors were corrected by a native speaker against the originals (5,833 sentences). A further 2,838 sentences are digital texts from local community authors, spell-checked by a native speaker. The last 101 sentences are German MASSIVE utterances, manually translated with the Braun and Mangold (1984) dictionary, with entities replaced by local place names and personal names and numerals spelled out in dialect orthography so speakers pronounce them consistently.

Speakers and recording (§3.2, §3.3). Nine volunteers (4 female, 5 male; ages from 18-25 to 61-65) recruited by convenience sampling, all native speakers of a Rhine Franconian dialect or a closely related regional dialect with long-term exposure. Self-reported labels vary: five say Saarländisch, and one each says Moselle Franconian, Rhine Franconian, Rhine-Moselle border dialect, and Platt. Recordings were made in a soundproof booth at 44.1 kHz with SpeechRecorder, one sentence on screen at a time, about 200-300 sentences per roughly one-hour session. Rather than sampling sentences independently, the concatenated text was cut into chunks of 100 consecutive sentences and whole chunks were randomly assigned to sessions. The authors argue this keeps topical continuity and supports more natural prosody and consistent pronunciation of recurring words, and that it makes the book subset usable for long-context TTS. They test none of this.

Analysis (§4). The authors run the Epitran multilingual G2P (Standard German mode) over the transcripts to estimate phoneme coverage, compute per-speaker F0 and speaking-rate statistics, and estimate recording quality with WADA-SNR.

Key Results

All results describe the corpus. None comes from a speech-generation experiment.

  • Size (Table 2, Table 3): 4,871 recorded utterances, 5:54:48 total, mean 4.37 s per utterance (min 1.59 s, max 14.02 s). Only a subset of the 8,772 collected sentences was recorded. Per-speaker contributions are uneven: two speakers (P02 with 1,075 sentences and P03 with 899) together provide over a third of the audio, while four speakers have about 300 sentences each.
  • Speaker acoustics (§4.3, Table 3): mean F0 is 115-171 Hz for male speakers and 201-251 Hz for female speakers, and speaking rate ranges from 1.50 to 2.42 words per second.
  • Phoneme coverage (§4.2): Epitran output contains 38 distinct phonemes. /d/ is the most frequent (13,963 occurrences), which the authors attribute to dialectal voicing of Standard German /t/. Front rounded vowels are rare (/ø/ 4, /œ/ 6, /ʏ/ 13, /y/ 26) and the authors treat them as Standard German loanword phonemes. Section 5.3 shows Epitran’s output is unreliable on this text, so these counts are rough estimates.
  • Audio quality (§4.4, Figure 2): WADA-SNR values range from 20 to 99 for most samples, and the authors read this as clean audio. No intelligibility, naturalness, or listener-based quality assessment is reported.
  • G2P failure analysis (§5.3, Table 4): Standard-German Epitran systematically lengthens word-final ‘e’ to /e:/ instead of schwa, reads dialect double vowels (e.g. ‘buuch’, ‘ääner’) as two long vowels, segments dialect ‘dd’ as /td/, and emits symbols that are impossible in German for the ‘á’ and ‘ò’ characters. Table 4 gives native-speaker corrected transcriptions for eight example words, and the authors label these corrections as estimates.

Novelty Assessment

The novelty is the resource itself: as far as the authors know, it is the first multi-speaker, studio-quality read-speech corpus for the Saarbrücken dialect beyond a small Bible MMS subset. The construction methods (OCR plus native-speaker correction, community-sourced text, localized MASSIVE translation, chunked read-speech recording) are standard corpus-building practice. The G2P error analysis is qualitative and covers only one off-the-shelf tool. The paper does not train, fine-tune, or evaluate any TTS or ASR system on the corpus, so its claimed suitability for “training and evaluating modern TTS systems” (§2) is a design intention. The paper explicitly defers zero-shot and fine-tuned TTS experiments with XTTS and ZMM-TTS to future work (§6, §7).

Field Significance

Low — this paper adds a small dialect read-speech corpus aimed at TTS for an underrepresented German variety, along with a useful qualitative record of how Standard-German text front-ends fail on unstandardized dialect spelling. Its value to speech generation depends on downstream TTS experiments the paper does not run.

Claims

  • supports: Grapheme-to-phoneme tools built for a standard language transfer poorly to unstandardized dialect orthography, producing systematic rather than random pronunciation errors in a TTS front-end.

    Evidence: Standard-German Epitran applied to Saarbrücken dialect text assigns vowel length incorrectly (word-final ‘e’ as /e:/ instead of schwa, double vowels read as two long vowels), segments dialect ‘dd’ as /td/, and outputs symbols that are impossible in German for dialect diacritics ‘á’ and ‘ò’. Native-speaker estimated corrections differ from the Epitran output on all eight Table 4 example words. (§5.3, Table 4)

  • complicates: Dialect speech corpora built from available written sources inherit a written-register and genre bias that may skew the prosody a TTS model learns.

    Evidence: Saar-Voice text is drawn from poetry, prose, folktales, and translated assistant utterances. Poetry makes up 35.9% of the data, and the authors note its metrical and rhyming conventions may introduce prosodic patterns unlike natural spoken dialect. The corpus contains no spontaneous speech. (§1.4, §5.1, Table 1)

  • complicates: In non-standardized dialects, orthographic variation means lexical and vocabulary statistics overcount types, which complicates text normalization and lexicon construction for dialect TTS.

    Evidence: The dialect has no official spelling, so writers spell the same word several ways (‘gebbd, gebd, gäbbd, …’). With no dialect tokenizer, vocabulary is approximated by whitespace tokenization, and spelling variants of one lexical item count as distinct types in the 11,303 unique-word figure. (§1.2, §5.1, Table 1)

  • complicates: Multi-speaker dialect corpora recorded from community volunteers capture a spectrum of dialect realizations rather than one uniform variety, which adds variability that downstream speech models must absorb.

    Evidence: The nine speakers self-label five different dialect varieties, including Moselle Franconian and a Rhine-Moselle border dialect. Most describe their speech as a mix of dialect and Standard German, only one reports speaking near-exclusively dialect, and per-speaker data ranges from 300 to 1,075 sentences. (§3.2, §5.2, Table 3)

Limitations and Open Questions

Warning

No TTS (or any other model) is trained, fine-tuned, or evaluated on Saar-Voice. Its suitability for dialect TTS, zero-shot adaptation, or long-context synthesis is asserted but not demonstrated, and the authors defer all synthesis experiments to future work (§6, §7).

  • Scale and balance: about six hours from nine speakers, with uneven per-speaker coverage (about 18 minutes to 1 h 16 min each). This is small for training multi-speaker TTS from scratch and is mainly suited to adaptation experiments, which the paper does not run.
  • Speaker sampling: convenience sampling from people known to the team, and the speakers’ dialect backgrounds are heterogeneous (§3.2, §5.2). The corpus does not represent one controlled dialect point.
  • Phonetic annotation: the only phonetic layer comes from a Standard-German G2P that the paper itself shows is unreliable on this text (§5.3). No validated dialect lexicon or forced alignment is reported, so the phoneme-coverage figures in §4.2 are rough estimates.
  • Quality assessment: audio quality is judged only by WADA-SNR. No check of transcript-audio fidelity (e.g. reading errors or deviations from the prompt) is reported beyond the option to re-record during sessions.
  • Register: read speech of written texts only, with a large poetry share and no spontaneous speech (§5.1).
  • Release: the parsed footnote states the dataset is available on Hugging Face (UdS-LSV/Saar-Voice). No code is described.

Wiki Connections

  • XTTS — named as one of two multilingual zero-shot TTS models (with ZMM-TTS) whose pretraining includes German and which the authors plan to evaluate or fine-tune on Saar-Voice in future work.
  • MLS — cited as an existing Standard German TTS/ASR corpus whose lack of dialect coverage motivates Saar-Voice.