HUI German 51 Speakers Synthetic (Cleaned)
Multispeaker German TTS dataset created and published by @dida-80b.
The voices are derived from one reference recording per speaker from the HUI-Audio-Corpus-German. The released audio here is newly synthesized with Qwen3-TTS; it is not a repackaging of the original HUI audio corpus.
Dataset
- 51 synthetic speaker voices derived from one HUI reference recording per speaker
- ~102h synthetic audio
- Audio synthesized with Qwen3-TTS from German transcript text
- Clips segmented to max ~15s, 24 kHz mono WAV
Validation status
This release was checked locally after publication metadata repair:
- 30,858 referenced WAV files
- 0 missing referenced WAV files
- 0 unreferenced WAV files
- 51 speakers
- 24 kHz mono WAV throughout
If the Hugging Face Dataset Viewer is unavailable for this repository, use the files directly via train_list.txt, val_list.txt, speaker_map.txt, and the audio/ directory.
Splits
| Split | Samples |
|---|---|
| Train | 29,339 |
| Validation | 1,519 |
| Total | 30,858 |
(Initial release list: 30,859 — one outlier sample excluded, see below)
Release Notes
1. ?? Encoding Artifact → ʊɾ (8,162 samples fixed)
The initial published list files contained ?? artifacts where the IPA phoneme bigram ʊɾ was corrupted during the synthesis/export process. This affected 8,162 samples (7,804 train + 358 val), or ~26.4% of the dataset.
The fix is a deterministic substitution (s/??/ʊɾ/g), verified safe through context analysis — in 100% of cases, ʊɾ is the only plausible reconstruction.
Before: d??ç → After: dʊɾç („durch")
Before: fɛɾˈ??zaxt → After: fɛɾˈʊɾzaxt („verursacht")
Before: kˈ??ʦ → After: kˈʊɾʦ („kurz")
2. Outlier sample excluded
One sample was removed from the train set:
Robert_Steiner_syn_0096.wav— duration 42.1s, flagged as extreme outlier (>30s threshold). This is the only sample excluded. The WAV file still exists in the audio directory but is not referenced in any list.
3. Wf token preserved
The token Wf (encoding the phoneme sequence /aʊf/, as in German "auf") appears in 11,160 samples and is not an error. It is a legitimate token in the Qwen3-TTS phoneme scheme.
4. Audio cleanup
All 30,858 clips were automatically cleaned:
- Breath removal: 26,461 clips (85.8%) — audible breathing sounds removed cleanly.
- Start-of-clip cleanup: 6,024 clips (19.5%) — leading silence or breath noise trimmed.
- End-of-clip repair: 764 clips (2.5%) — truncated waveforms extended to zero-crossing to prevent clicks.
Original unprocessed audio is not included in this release.
Format
audio/<speaker_name>/wavs/<filename>.wav
train_list.txt
val_list.txt
speaker_map.txt
README.md
train_list.txt/val_list.txt: Pipe-separated format:<wav_path>|<phoneme_transcript>|<speaker_idx>- WAV paths use the form
audio/<speaker_name>/wavs/<filename>.wav - Phoneme transcript uses IPA-based Qwen3-TTS phoneme scheme
speaker_map.txt: Tab-separated mapping of<speaker_idx> <speaker_name>- Audio: 24 kHz mono WAV
Source Relation
| Aspect | Source / initial release | This dataset version |
|---|---|---|
| Speaker source | HUI-Audio-Corpus-German reference recordings | One reference recording per speaker used to derive synthetic voices |
| Released audio | Not original HUI corpus audio | Newly synthesized Qwen3-TTS audio |
?? artifacts | Present in initial list files (8,162 samples) | Fixed — substituted with ʊɾ |
| Sample count | Initial list: 30,859 | 30,858 |
Robert_Steiner_syn_0096.wav | Initially referenced in train set | Excluded (outlier, 42.1s) |
Wf token | Present | Preserved (legitimate) |
| Breath sounds | Present in original synthesis | Removed — 26,461 clips (85.8%) |
| Start-of-clip noise | Present | Removed — 6,024 clips (19.5%) |
| Truncated endings | Present | Repaired — 764 clips (2.5%) |
Credits
- Dataset creation, synthesis release, cleaning, verification, and publication: @dida-80b (May 2026)
- Reference speaker recordings: HUI-Audio-Corpus-German (Florian Lux et al., Hochschule Hof)
- Transcript / pipeline basis: @semidark, including work published via semidark/hokuspokus-qwen3-tts-hybrid
- Audio synthesis: Qwen3-TTS
If this cleaned release is useful for your work, please cite or link this dataset page and consider leaving a like on Hugging Face. It helps make maintenance and future dataset cleanup visible.
License
No formal license declared. The HUI-Audio-Corpus-German is based on LibriVox recordings (Public Domain in the USA; legal status varies by country). The HUI team (Hochschule Hof) requests attribution in the spirit of CC-BY-SA 4.0 but does not legally enforce it.
Synthetic audio in this dataset was generated with Qwen3-TTS from derived speaker references and German transcript text; no original HUI corpus audio is included as released training audio.