aman4014/translated-german-english-asr

---
dataset_info:
features:
- name: audio
dtype: audio
- name: transcription
dtype: string
- name: translation
dtype: string
splits:
- name: train_el_tts
num_bytes: 702196490.0
num_examples: 495
- name: train_mls_0
num_bytes: 191221465088.0
num_examples: 406042
- name: train_tuda_0
num_bytes: 13599196098.12
num_examples: 42812
- name: train_cv19_0
num_bytes: 245838579828.326
num_examples: 546561
- name: train_emilia_yodas0
num_bytes: 955252797694.16
num_examples: 1980468
- name: train_eurospeech
num_bytes: 241659687916.69
num_examples: 502515
- name: train_de_DE_kerstin
num_bytes: 139989365.876
num_examples: 1374
- name: TV_2021.02_Neutral
num_bytes: 7254720000.0
num_examples: 22671
- name: TV_2021.06_Emotional
num_bytes: 646400000.0
num_examples: 2020
- name: TV_2022.10_Neutral
num_bytes: 3984320000.0
num_examples: 12451
- name: TV_2023.09_Hessisch
num_bytes: 673920000.0
num_examples: 2106
- name: train_mozilla_english_asr
num_bytes: 46096297644
num_examples: 1147812
download_size: 1707069570125.172
dataset_size: 1707069570125.172
configs:
- config_name: default
data_files:
- split: train_el_tts
path: data/train_el_tts-*
- split: train_mls_0
path: data/train_mls_0-*
- split: train_tuda_0
path: data/train_tuda_0-*
- split: train_cv19_0
path: data/train_cv19_0-*
- split: train_emilia_yodas0
path: data/train_emilia_yodas0-*
- split: train_eurospeech
path: data/train_eurospeech-*
- split: train_de_DE_kerstin
path: data/train_de_DE_kerstin-*
- split: TV_2021.02_Neutral
path: data/TV_2021.02_Neutral-*
- split: TV_2021.06_Emotional
path: data/TV_2021.06_Emotional-*
- split: TV_2022.10_Neutral
path: data/TV_2022.10_Neutral-*
- split: TV_2023.09_Hessisch
path: data/TV_2023.09_Hessisch-*
- split: train_mozilla_english_asr
path: data/train_mozilla_english_asr-*
language:
- de
- en
task_categories:
- automatic-speech-recognition
- translation
tags:
- german
- speech
- asr
- tts
- translation
- multilingual
pretty_name: Translated German-English ASR Dataset
---

# Translated German-English ASR Dataset

A large-scale, multi-source German speech dataset with paired English translations, designed for training and evaluating German **Automatic Speech Recognition (ASR)**, **Speech Translation**, and **Text-to-Speech (TTS)** systems. This dataset is a curated mixture of well-established open-source German and multilingual speech corpora, all unified under a common schema with German audio, original German transcriptions, and English translations.

<table style="border-collapse: collapse; border: none;">
<tr style="border: none;">
<td style="border: none; padding: 0 20px;">
<a href="https://hpi.de/ki-servicezentrum/">
<img src="https://docs.sc.hpi.de/attachments/aisc/aisc-logo.png" alt="KI-Servicezentrum Berlin-Brandenburg" style="height: 60px; width: auto;">
</a>
</td>
<td style="border: none; padding: 0 20px;">
<a href="https://www.bmftr.bund.de">
<img src="https://docs.sc.hpi.de/attachments/aisc/bmftr.jpg" alt="Gefoerdert durch BMFTR" style="height: 60px; width: auto;">
</a>
</td>
</tr>
</table>

---

## Dataset Summary

| Property | Value |
|---|---|
| **Primary Language** | German (de) |
| **Translation Language** | English (en) |
| **Total Examples** | 4,667,327 |
| **Total Size (disk)** | ~1.71 TB (download) / ~1.71 TB (uncompressed) |
| **Audio Format** | Variable (WAV/FLAC/MP3, 16kHz-44.1kHz) |
| **Tasks** | ASR, Speech Translation, TTS |

---

## Dataset Structure

### Features

Each example contains the following fields:

| Field | Type | Description |
|---|---|---|
| `audio` | `Audio` | The audio file, automatically decoded and resampled on access |
| `transcription` | `string` | Original German transcription of the spoken audio |
| `translation` | `string` | English translation of the German transcription |

### Splits

The dataset is organized into 12 splits, each sourced from a distinct German speech corpus:

| Split | Source Corpus | Examples | Approx. Size | License |
|---|---|---|---|---|
| `train_el_tts` | Custom TTS (Greek-Letters / Eliza-style German TTS) | 495 | ~670 MB | Contact provider |
| `train_mls_0` | [Multilingual LibriSpeech (MLS) - German](https://huggingface.co/datasets/facebook/multilingual_librispeech) | 406,042 | ~182 GB | CC BY 4.0 |
| `train_tuda_0` | [Tuda-De (TU Darmstadt German ASR)](https://huggingface.co/datasets/uhhlt/Tuda-De) | 42,812 | ~13 GB | CC BY 4.0 |
| `train_cv19_0` | [Mozilla Common Voice 19 - German](https://commonvoice.mozilla.org/de/datasets) | 546,561 | ~234 GB | CC0 1.0 |
| `train_emilia_yodas0` | [Emilia-YODAS (German subset)](https://huggingface.co/datasets/amphion/Emilia-Dataset) | 1,980,468 | ~911 GB | CC BY 4.0 |
| `train_eurospeech` | [EuroSpeech - German Parliament](https://huggingface.co/datasets/disco-eth/EuroSpeech) | 502,515 | ~230 GB | Per-parliament (see below) |
| `train_de_DE_kerstin` | [M-AILABS - de_DE_kerstin](https://www.caito.de/2019/01/03/the-m-ailabs-speech-dataset/) | 1,374 | ~134 MB | M-AILABS BSD-3-Clause style |
| `TV_2021.02_Neutral` | [Thorsten-Voice 2021.02 - Neutral](https://huggingface.co/datasets/Thorsten-Voice/TV-44kHz-Full) | 22,671 | ~6.9 GB | CC0 1.0 |
| `TV_2021.06_Emotional` | [Thorsten-Voice 2021.06 - Emotional](https://huggingface.co/datasets/Thorsten-Voice/TV-44kHz-Full) | 2,020 | ~617 MB | CC0 1.0 |
| `TV_2022.10_Neutral` | [Thorsten-Voice 2022.10 - Neutral](https://huggingface.co/datasets/Thorsten-Voice/TV-44kHz-Full) | 12,451 | ~3.8 GB | CC0 1.0 |
| `TV_2023.09_Hessisch` | [Thorsten-Voice 2023.09 - Hessisch Dialect](https://huggingface.co/datasets/Thorsten-Voice/TV-44kHz-Full) | 2,106 | ~642 MB | CC0 1.0 |
| `train_mozilla_english_asr` | Mozilla Common Voice 25 - English (translated) | 1,147,812 | ~42.9 GB | CC0 1.0 |

**Total: ~4,667,327 examples · ~1.71 TB**

---

## Usage

### Load a Specific Split

```python
from datasets import load_dataset

# Load a single split
dataset = load_dataset("aman4014/translated-german-english-asr", split="train_mls_0")
print(dataset[0])
```

### Load All Splits

```python
from datasets import load_dataset

dataset = load_dataset("aman4014/translated-german-english-asr")
print(dataset.keys())
# dict_keys(['train_el_tts', 'train_mls_0', 'train_tuda_0',
# 'train_cv19_0', 'train_emilia_yodas0', 'train_eurospeech',
# 'train_de_DE_kerstin', 'TV_2021.02_Neutral', 'TV_2021.06_Emotional',
# 'TV_2022.10_Neutral', 'TV_2023.09_Hessisch', 'train_mozilla_english_asr'])
```

### Streaming (Recommended for Large Splits)

```python
from datasets import load_dataset

# Stream large splits to avoid downloading everything at once
dataset = load_dataset(
"aman4014/translated-german-english-asr",
split="train_emilia_yodas0",
streaming=True
)

for example in dataset.take(5):
print(example["transcription"])
print(example["translation"])
```

### Access Audio

```python
from datasets import load_dataset

dataset = load_dataset("aman4014/translated-german-english-asr", split="train_cv19_0")

# Audio is decoded on access
sample = dataset[0]
audio_array = sample["audio"]["array"] # numpy array
sampling_rate = sample["audio"]["sampling_rate"] # e.g. 16000
transcription = sample["transcription"] # German text
translation = sample["translation"] # English text
```

---

## Source Datasets and Descriptions

### Multilingual LibriSpeech (MLS) - `train_mls_0`

MLS is a large-scale multilingual corpus derived from LibriVox audiobooks, covering 8 languages including German. The German subset contains ~1,000 hours of read-speech data from public-domain books. Produced by Facebook AI Research (Meta).

- **Paper:** [MLS: A Large-Scale Multilingual Dataset for Speech Research](https://arxiv.org/abs/2012.03411) (Pratap et al., 2020)
- **Source:** [facebook/multilingual_librispeech](https://huggingface.co/datasets/facebook/multilingual_librispeech)
- **License:** CC BY 4.0
- **Style:** Read speech (audiobooks)

### Emilia-YODAS - `train_emilia_yodas0`

Emilia-YODAS is a large-scale multilingual speech dataset processed via the Emilia-Pipe pipeline. This subset is part of the larger Emilia-Large release (~216,000 hours total) and is released under the permissive CC BY 4.0 license.

- **Paper:** [Emilia: A Large-Scale, Extensive, Multilingual, and Diverse Dataset](https://arxiv.org/abs/2501.15907) (He et al., 2025)
- **Source:** [amphion/Emilia-Dataset](https://huggingface.co/datasets/amphion/Emilia-Dataset)
- **License:** CC BY 4.0
- **Style:** In-the-wild / spontaneous speech

### Tuda-De - `train_tuda_0`

The Tuda-De corpus is a German read-speech dataset recorded at TU Darmstadt using multiple microphones (Kinect, Realtek, Headset). Speakers read sentences from German Wikipedia, the Europarl corpus, and web-crawled text. It is one of the foundational open-source German ASR datasets.

- **Paper:** [Open Source German Distant Speech Recognition: Corpus and Acoustic Model](https://doi.org/10.1007/978-3-319-24033-6_54) (Radeck-Arneth et al., 2015)
- **Source:** [uhhlt/Tuda-De](https://huggingface.co/datasets/uhhlt/Tuda-De)
- **License:** CC BY 4.0
- **Style:** Read speech, multi-microphone, controlled environment

### Common Voice 19 - `train_cv19_0`

Mozilla Common Voice is a massively multilingual, crowd-sourced speech corpus. Volunteers record text prompts and validate each other's recordings, resulting in diverse speech with varied accents, ages, and genders. Version 19 covers 129+ languages.

- **Paper:** [Common Voice: A Massively-Multilingual Speech Corpus](https://arxiv.org/abs/1912.06670) (Ardila et al., 2020)
- **Source:** [mozilla-foundation/common_voice_19_0](https://commonvoice.mozilla.org/de/datasets)
- **License:** CC0 1.0 (Public Domain)
- **Style:** Crowd-sourced read speech, diverse speakers

### EuroSpeech - `train_eurospeech`

EuroSpeech is a large-scale multilingual corpus of parliamentary speech from 22 European nations, aligned using a novel two-stage dynamic algorithm. The German subset is sourced from Bundestag/Bundesrat sessions. Licensing reflects the open-access policies of each national parliament; parliamentary speech in most European jurisdictions is released for public use.

- **Paper:** [EuroSpeech: A Multilingual Speech Corpus](https://arxiv.org/abs/2510.00514) (Pfisterer et al., 2025)
- **Source:** [disco-eth/EuroSpeech](https://huggingface.co/datasets/disco-eth/EuroSpeech)
- **License:** Per-country parliamentary open-access terms (see dataset card for full breakdown)
- **Style:** Parliamentary / formal speech

### M-AILABS - de_DE_kerstin - `train_de_DE_kerstin`

The M-AILABS Speech Dataset is a multi-language TTS/ASR corpus based on LibriVox public domain audiobooks and Project Gutenberg texts. The `de_DE_kerstin` split corresponds to a single German female speaker ("Kerstin") reading audiobook passages.

- **Source:** [M-AILABS Speech Dataset](https://www.caito.de/2019/01/03/the-m-ailabs-speech-dataset/)
- **License:** M-AILABS BSD-3-Clause style license (attribution required, no endorsement)
- **Style:** Read speech (audiobooks), single speaker

### Thorsten-Voice - `TV_2021.02_Neutral`, `TV_2021.06_Emotional`, `TV_2022.10_Neutral`, `TV_2023.09_Hessisch`

Thorsten-Voice is a freely contributed German TTS voice dataset by Thorsten Muller, a single male native German speaker. It encompasses multiple recording sessions covering neutral speech, emotional speech (angry, disgusted, amused, drunk, surprised, sleepy, whisper), an updated neutral session, and a Hessian dialect (Hessisch) session. All recordings are at 44.1kHz stereo and are released under the completely unrestricted CC0 public domain license.

- **Source:** [Thorsten-Voice/TV-44kHz-Full](https://huggingface.co/datasets/Thorsten-Voice/TV-44kHz-Full)
- **License:** CC0 1.0 (Public Domain)
- **Style:** TTS-quality read speech, single male native German speaker

| Subset | Recording Session | Style |
|---|---|---|
| `TV_2021.02_Neutral` | Feb 2021 | Neutral, clear |
| `TV_2021.06_Emotional` | Jun 2021 | Emotional (7 emotion categories) |
| `TV_2022.10_Neutral` | Oct 2022 | Neutral, high quality, LJSpeech-compatible |
| `TV_2023.09_Hessisch` | Sep 2023 | Hessian regional dialect |

### Mozilla Common Voice 25 - English - `train_mozilla_english_asr`

The Mozilla Common Voice English split uses validated English clips with English transcriptions. This dataset uses the English audio and translates the transcription into German to keep the same schema as the German corpora.

- **Source:** [mozilla-foundation/common_voice_25_0](https://commonvoice.mozilla.org/en/datasets)
- **License:** CC0 1.0 (Public Domain)
- **Style:** Crowd-sourced read speech, diverse speakers

---

## Dataset Sources and Licensing

This dataset is a mixture of several German and multilingual speech datasets. For each dataset, the license of the **original author** applies. Please consult the linked sources for detailed licensing information and terms of use.

| Split | Source Dataset | License | Commercial Use | Link |
|---|---|---|---|---|
| `train_el_tts` | Custom TTS | Unknown - contact provider | Unknown | N/A |
| `train_mls_0` | Multilingual LibriSpeech (MLS) | **CC BY 4.0** | Yes | [openslr.org/94](https://www.openslr.org/94/) |
| `train_tuda_0` | Tuda-De | **CC BY 4.0** | Yes | [uhhlt/Tuda-De](https://huggingface.co/datasets/uhhlt/Tuda-De) |
| `train_cv19_0` | Mozilla Common Voice 19 | **CC0 1.0** | Yes | [commonvoice.mozilla.org](https://commonvoice.mozilla.org/de/datasets) |
| `train_emilia_yodas0` | Emilia-YODAS | **CC BY 4.0** | Yes | [amphion/Emilia-Dataset](https://huggingface.co/datasets/amphion/Emilia-Dataset) |
| `train_eurospeech` | EuroSpeech | Per-parliament open access | Verify per country | [disco-eth/EuroSpeech](https://huggingface.co/datasets/disco-eth/EuroSpeech) |
| `train_de_DE_kerstin` | M-AILABS | M-AILABS BSD-3-Clause | Yes (with attribution) | [caito.de](https://www.caito.de/2019/01/03/the-m-ailabs-speech-dataset/) |
| `TV_2021.02_Neutral` | Thorsten-Voice | **CC0 1.0** | Yes | [Thorsten-Voice/TV-44kHz-Full](https://huggingface.co/datasets/Thorsten-Voice/TV-44kHz-Full) |
| `TV_2021.06_Emotional` | Thorsten-Voice | **CC0 1.0** | Yes | [Thorsten-Voice/TV-44kHz-Full](https://huggingface.co/datasets/Thorsten-Voice/TV-44kHz-Full) |
| `TV_2022.10_Neutral` | Thorsten-Voice | **CC0 1.0** | Yes | [Thorsten-Voice/TV-44kHz-Full](https://huggingface.co/datasets/Thorsten-Voice/TV-44kHz-Full) |
| `TV_2023.09_Hessisch` | Thorsten-Voice | **CC0 1.0** | Yes | [Thorsten-Voice/TV-44kHz-Full](https://huggingface.co/datasets/Thorsten-Voice/TV-44kHz-Full) |
| `train_mozilla_english_asr` | Mozilla Common Voice 25 - English (translated) | **CC0 1.0** | Yes | [commonvoice.mozilla.org](https://commonvoice.mozilla.org/en/datasets) |

> **Note:** If a dataset does not have a public source listed, please contact the dataset provider or refer to your data distributor for licensing details.

### License Summary

- **CC0 1.0 (Public Domain):** No restrictions. Can be used commercially, modified, and redistributed without attribution. Applies to Common Voice and all Thorsten-Voice splits.
- **CC BY 4.0:** Free to use commercially and non-commercially, with attribution. Applies to MLS, Tuda-De, and Emilia-YODAS.
- **M-AILABS BSD-3-Clause:** Similar to BSD-3. Commercial use permitted with attribution; no endorsement of products derived from the data.
- **EuroSpeech / Parliamentary:** Licensed under the open-access terms of each respective national parliament. Generally free for research use; commercial use should be verified per country.

---

## Licensing Disclaimer

The use of this dataset and any derived models must comply with the licenses of the **original underlying datasets**. The most restrictive license in any given use case applies:

- **For fully open / commercial use:** Verify EuroSpeech country-level terms.
- **For non-commercial research:** All splits are generally usable, subject to attribution where required.
- **For the `train_el_tts` split:** License information is unknown. Contact the dataset provider before use.

The dataset creators make no representations or warranties regarding these datasets, including warranties of non-infringement or fitness for a particular purpose. The dataset creators do not claim any rights to the datasets themselves - all rights remain with the original data owners.

**Always verify the license terms before using this data or any models trained on it for commercial or research purposes.**

---

## Acknowledgements

This dataset would not have been possible without the contributions of the following organizations and open-source communities:

* **Mozilla Common Voice**: For providing a massively multilingual, freely available crowd-sourced speech corpus
* **Facebook AI Research (Meta)**: For the Multilingual LibriSpeech dataset
* **Amphion / Emilia Team**: For the Emilia-YODAS large-scale multilingual speech dataset
* **TU Darmstadt**: For the Tuda-De German ASR corpus
* **disco-eth / EuroSpeech Team**: For the multilingual European parliamentary speech corpus
* **Thorsten Muller**: For the freely contributed Thorsten-Voice German TTS dataset
* **M-AILABS**: For the multilingual audiobook-based speech corpus
* **AI Service Center Berlin-Brandenburg (KI-Servicezentrum)**: For supporting this work

<table style="border-collapse: collapse; border: none;">
<tr style="border: none;">
<td style="border: none; padding: 0 20px;">
<a href="https://hpi.de/ki-servicezentrum/">
<img src="https://docs.sc.hpi.de/attachments/aisc/aisc-logo.png" alt="KI-Servicezentrum Berlin-Brandenburg" style="height: 60px; width: auto;">
</a>
</td>
<td style="border: none; padding: 0 20px;">
<a href="https://www.bmftr.bund.de">
<img src="https://docs.sc.hpi.de/attachments/aisc/bmftr.jpg" alt="Gefoerdert durch BMFTR" style="height: 60px; width: auto;">
</a>
</td>
</tr>
</table>

**Funding Notice**

Das zugrunde liegende Vorhaben wurde mit Mitteln des Bundesministeriums fur Forschung, Technologie und Raumfahrt unter dem Foerderkennzeichen "KI-Servicezentrum Berlin-Brandenburg" 16IS22092 gefoerdert. Die Verantwortung fur den Inhalt dieser Veroeffentlichung liegt beim Autor.

_This project was funded by the German Federal Ministry of Research, Technology and Space under the funding code "KI-Servicezentrum Berlin-Brandenburg" 16IS22092. Responsibility for the content of this publication remains with the author._

---

## Citation

If you use this dataset, please cite the original source datasets as appropriate:

**Multilingual LibriSpeech:**
```bibtex
@article{Pratap2020MLSAL,
title={MLS: A Large-Scale Multilingual Dataset for Speech Research},
author={Vineel Pratap and Qiantong Xu and Anuroop Sriram and Gabriel Synnaeve and Ronan Collobert},
journal={ArXiv},
year={2020},
volume={abs/2012.03411}
}
```

**Emilia-YODAS:**
```bibtex
@inproceedings{emilia,
author={He, Haorui and Shang, Zengqiang and Wang, Chaoren and Li, Xuyuan and Gu, Yicheng
and Hua, Hua and Liu, Liwei and Yang, Chen and Li, Jiaqi and Shi, Peiyang
and Wang, Yuancheng and Chen, Kai and Zhang, Pengyuan and Wu, Zhizheng},
title={Emilia: An Extensive, Multilingual, and Diverse Speech Dataset for Large-Scale Speech Generation},
booktitle={Proc. of SLT},
year={2024}
}
```

**Tuda-De:**
```bibtex
@InProceedings{Radeck-Arneth2015,
author="Radeck-Arneth, Stephan and Milde, Benjamin and Lange, Arvid and Gouvea, Evandro
and Radomski, Stefan and Muhlhauser, Max and Biemann, Chris",
title="Open Source German Distant Speech Recognition: Corpus and Acoustic Model",
booktitle="Text, Speech, and Dialogue",
year="2015",
publisher="Springer International Publishing",
pages="480--488",
doi="10.1007/978-3-319-24033-6_54"
}
```

**Mozilla Common Voice:**
```bibtex
@inproceedings{commonvoice:2020,
author = {Ardila, R. and Branson, M. and Davis, K. and Henretty, M. and Kohler, M.
and Meyer, J. and Morais, R. and Saunders, L. and Tyers, F. M. and Weber, G.},
title = {Common Voice: A Massively-Multilingual Speech Corpus},
booktitle = {Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020)},
pages = {4211--4215},
year = 2020
}
```

**EuroSpeech:**
```bibtex
@article{pfisterer2025eurospeech,
title={EuroSpeech: A Multilingual Speech Corpus},
author={Samuel Pfisterer and Florian Grotschla and Luca Lanzendorfer and Florian Yan and Roger Wattenhofer},
year={2025}
}
```

**Thorsten-Voice:**
```bibtex
@misc{thorsten_muller_2024,
author = {{Thorsten Muller}},
title = {TV-44kHz-Full},
year = 2024,
url = {https://huggingface.co/datasets/Thorsten-Voice/TV-44kHz-Full},
doi = {10.57967/hf/3290},
publisher = {Hugging Face}
}
```

**M-AILABS:**
```bibtex
@misc{MAILABS_2017,
author = {Solak, I. Celeste Aurora and Naumov, Dima},
title = {The M-AILABS Speech Dataset},
year = {2017},
howpublished = {\url{https://github.com/i-celeste-aurora/m-ailabs-dataset}}
}
```