German European Parliament plenary speech paired with a transcript, acoustic measurements, discrete German descriptor tags and a free-text German description of the voice and recording conditions. The dataset is intended for training description-conditioned TTS models such as Parler-TTS.

The data was built as part of research work. It is a random subset of the German unlabelled part of VoxPopuli (release v1, from the VoxPopuli GitHub repository), not the transcribed subset on the Hub (facebook/voxpopuli), which it does not overlap (see below).

Most of the audio is simultaneous interpretation. VoxPopuli's German unlabelled data is the Parliament's German audio channel: when a member speaks German it carries the original speech, otherwise the German interpreters. The German original speech is largely already in facebook/voxpopuli and was removed, so the large majority of this dataset (roughly nine in ten clips) is interpretation by the German booth -- professional speakers, studio microphones, but interpretation prosody, and a comparatively small number of recurring voices (see the speaker mapping below).

Splits

Split	Content	`text_description`
`female`	clips predicted female (`weiblich`), up to 860 hours	German voice description
`male`	clips predicted male (`männlich`), up to 860 hours	German voice description
`extra`	every clip beyond its gender's 860 hours	`null`

Clips are assigned in a fixed random order: each goes to the split of its predicted gender until that split holds 860 hours, every later clip of that gender goes to extra. The German channel is about 70% female, so female fills first and extra is almost entirely female. The split is decided by a gender classifier (see Columns), not by human labels, so a small share of clips is in the wrong split.

Sources and attribution

Source	Where	License
VoxPopuli v1 unlabelled, German (`--subset de`: plenary sessions 2009-2020)	facebookresearch/voxpopuli, `dl.fbaipublicfiles.com/voxpopuli/audios/de_<year>.tar`	CC0 1.0 (see also the European Parliament's legal notice for the raw data)
MOSEL transcripts (Whisper large-v3)	`FBK-MT/mosel`, `de/voxpopuli.tsv`, revision `40d224f`	CC BY 4.0

Because the MOSEL transcripts are CC BY 4.0, this derived dataset is released under CC BY 4.0. Please cite VoxPopuli and MOSEL:

@inproceedings{wang-etal-2021-voxpopuli,
    title = "{V}ox{P}opuli: A Large-Scale Multilingual Speech Corpus for Representation Learning, Semi-Supervised Learning and Interpretation",
    author = "Wang, Changhan  and Riviere, Morgane  and Lee, Ann  and Wu, Anne  and Talnikar, Chaitanya  and Haziza, Daniel  and Williamson, Mary  and Pino, Juan  and Dupoux, Emmanuel",
    booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)",
    month = aug,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.acl-long.80",
    pages = "993--1003",
}

@inproceedings{mosel,
  title = {{MOSEL: 950,000 Hours of Speech Data for Open-Source Speech Foundation Model Training on EU Languages}},
  author = {Marco Gaido and Sara Papi and Luisa Bentivogli and Alessio Brutti and Mauro Cettolo and Roberto Gretter and Marco Matassoni and Mohamed Nabih and Matteo Negri},
  booktitle = "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing",
  month = nov,
  year = "2024",
  address = "Miami, United States",
  publisher = "Association for Computational Linguistics",
}

No overlap with `facebook/voxpopuli`

facebook/voxpopuli (config de) was cut by VoxPopuli from the same plenary sessions. At the time stamps of its segments, the German channel used here contains the identical recording (sample-level correlation 0.93-0.97 after resampling, time offset under 0.4 s). Every unlabelled segment that comes within 2 seconds of any facebook/voxpopuli de segment (train, validation and test, taken from VoxPopuli's asr_de.tsv.gz) in the same session was removed (461 h), so no audio of that dataset appears here. leonhard-behr/voxpopuli-mls-de-descriptions, which used facebook/voxpopuli, therefore does not overlap this dataset either.

Transcripts and their quality

Transcripts are MOSEL's Whisper large-v3 output (with casing, punctuation and digits). Their quality was measured on the segments that coincide with human-derived facebook/voxpopuli transcripts (12,705 segments, 104 hours of German original speech; those segments are then excluded from the dataset): WER 12.2% overall, 8.2% after the filters below. On interpretation, where no human transcript exists, MOSEL agrees noticeably less with an independent ASR model than on original speech, so expect somewhat higher error rates there.

A segment is kept only if:

MOSEL's language ID is German and none of MOSEL's hallucination flags is set;
the transcript has 1.0-3.5 words per second (outside that range the transcript has lost most of the speech or loops: WER above 90%);
an independent CTC model (jonatasgrosman/wav2vec2-large-xlsr-53-german) agrees with it: character-level disagreement at most 0.4, and MOSEL's text covers at least 75% of the characters the CTC model hears (catches wrong-language audio, counting loops, and Whisper skipping part of the speech, typically a segment's opening clause).

asr_disagreement and asr_coverage are kept as columns for stricter filtering.

Columns

Column	Description
`audio`	The segment as 16 kHz 16-bit FLAC, cut sample-exactly as VoxPopuli's own segmentation does (15-30 s)
`text`	MOSEL transcript (Whisper large-v3), whitespace-normalised
`path`	Path of the exported clip during processing
`vp_id`	VoxPopuli segment id `<session event>_<segment number>`
`vp_start`, `vp_end`	Segment start and end in the session recording (seconds)
`asr_disagreement`, `asr_coverage`	Agreement of `text` with the independent CTC transcript (see above)
`phonemes`	IPA phonemization of `text` (`g2p`, `deu` -> `deu-ipa`)
`speaking_rate`	Phoneme units per second of detected speech
`utterance_pitch_mean`, `utterance_pitch_std`	Pitch statistics in Hz (PENN, FCNF0++ checkpoint, `local_expected_value` decoding, 10 ms hop, 30–1000 Hz)
`snr`, `c50`	Signal-to-noise ratio and C50 reverberation estimate (Brouhaha, `ylacombe/brouhaha-best`)
`stoi`, `si-sdr`, `pesq`	Reference-free quality estimates (torchaudio `SQUIM_OBJECTIVE`)
`gender`	Predicted speaker gender (`alefiury/wav2vec2-large-xlsr-53-gender-recognition-librispeech`); decides the split
`pitch`, `speaking_rate_bin`, `noise`, `reverberation`, `speech_monotony`, `sdr_noise`, `pesq_speech_quality`	German descriptor tags obtained by binning the measurements above
`text_description`	Free-text German description of voice and recording, generated from the tags (`null` in `extra`)

Speakers

VoxPopuli's unlabelled data has no speaker labels. Pseudo speaker ids (ECAPA embeddings, clustered within and then across sessions) are published separately in leonhard-behr/voxpopuli-v1-de-speakers, keyed by path and vp_id.

How it was made

The annotation follows the approach of Data-Speech (Lyth & King, 2024, Natural language guidance of high-fidelity text-to-speech with synthetic annotations), adapted for German, with the same pipeline and settings as leonhard-behr/voxpopuli-mls-de-descriptions. The descriptor bins are reused unchanged from that dataset, so a tag means the same measured range in both datasets and they can be combined for training.

Every 2,000-row shard was checked automatically before upload: rows, transcripts and embedded audio match the source byte for byte, tags recompute exactly from the measurements, split assignments recompute exactly, and per-stage alignment checks confirm that measurements belong to the right clips.

Limitations

Mostly simultaneous interpretation (see above), from a limited set of recurring interpreters.
Transcripts are automatic (Whisper large-v3); see the quality figures above. text_description is LLM-generated and is not always faithful to the tags; the tag columns are the authoritative annotation.
gender, and therefore the split, and all quality and acoustic columns are model predictions, not human labels.
The descriptor bins are relative to the VoxPopuli + MLS German distribution, not to any absolute scale.
For a small number of very short clips, Brouhaha returns no SNR/C50 (NaN); their noise and reverberation tags fall into the lowest bin and are unreliable.

promote_v1 (2026-10-06): +100 h per gender, noise-cell matched

Files data/female/promoted_v1-*.parquet and data/male/promoted_v1-*.parquet hold 12773 + 12835 clips (102.46 h women, 101.82 h men) taken from the extra split and given a description. They were chosen so that each gender follows the noise distribution (snr, c50, si-sdr, pesq; 5 bins each) of the VP v1 training target of the 2k draw, after the final filters and the held-out eval/test speakers. These clips are also still in extra (Hub files are never changed): drop the paths of the promoted files from extra if you use both. All files of revision 56bcde32 are unchanged.

Downloads last month: 1,234

Total file size:

253 GB

Request access to this dataset

Natural Language Voice Descriptions of the VoxPopuli v1 German (Unlabelled) Dataset