README.md

metadata

dataset_info:
  - config_name: TV-2021.02-Neutral
    description: >-
      Single german male speaker, neutral speech, very clear and precise, but
      low in speech flow.
    homepage: https://www.Thorsten-Voice.de
    license: CC0
    features:
      - name: audio
        dtype:
          audio:
            sampling_rate: 44100
      - name: id
        dtype: string
      - name: subset
        dtype: string
      - name: style
        dtype: string
      - name: text
        dtype: string
      - name: samplerate
        dtype: int32
      - name: durationSeconds
        dtype: float16
      - name: charsPerSecond
        dtype: float16
      - name: recording_year-month
        dtype: string
      - name: microphone
        dtype: string
      - name: speaker
        dtype: string
      - name: language
        dtype: string
      - name: comment
        dtype: string
    splits:
      - name: train
        num_bytes: 7290955038.594
        num_examples: 22671
    download_size: 6955484390
    dataset_size: 7290955038.594
  - config_name: TV-2021.06-Emotional
    description: >-
      Single german male speaker, all same phrases recorded in following
      emotions - Disgusted, Angry, Amused, Surprised, Sleepy, Drunk, Whispering
    homepage: https://www.Thorsten-Voice.de
    license: CC0
    features:
      - name: audio
        dtype:
          audio:
            sampling_rate: 44100
      - name: id
        dtype: string
      - name: subset
        dtype: string
      - name: style
        dtype: string
      - name: text
        dtype: string
      - name: samplerate
        dtype: int32
      - name: durationSeconds
        dtype: float16
      - name: charsPerSecond
        dtype: float16
      - name: recording_year-month
        dtype: string
      - name: microphone
        dtype: string
      - name: speaker
        dtype: string
      - name: language
        dtype: string
      - name: comment
        dtype: string
    splits:
      - name: train
        num_bytes: 793443429.88
        num_examples: 2020
    download_size: 748062212
    dataset_size: 793443429.88
  - config_name: TV-2022.10-Neutral
    description: >-
      Single german male speaker, neutral speech, very clear, high class
      quality, natural speech flow
    homepage: https://www.Thorsten-Voice.de
    license: CC0
    features:
      - name: audio
        dtype:
          audio:
            sampling_rate: 44100
      - name: id
        dtype: string
      - name: subset
        dtype: string
      - name: style
        dtype: string
      - name: text
        dtype: string
      - name: samplerate
        dtype: int32
      - name: durationSeconds
        dtype: float16
      - name: charsPerSecond
        dtype: float16
      - name: recording_year-month
        dtype: string
      - name: microphone
        dtype: string
      - name: speaker
        dtype: string
      - name: language
        dtype: string
      - name: comment
        dtype: string
    splits:
      - name: train
        num_bytes: 3559397446.61
        num_examples: 12451
    download_size: 3166262433
    dataset_size: 3559397446.61
  - config_name: TV-2023.09-Hessisch
    description: >-
      Single german male speaker, high class quality, natural speech flow,
      recorded in Hessisch and german dialect spoken in the middle of germany
    homepage: https://www.Thorsten-Voice.de
    license: CC0
    features:
      - name: audio
        dtype:
          audio:
            sampling_rate: 44100
      - name: id
        dtype: string
      - name: subset
        dtype: string
      - name: style
        dtype: string
      - name: text
        dtype: string
      - name: samplerate
        dtype: int32
      - name: durationSeconds
        dtype: float16
      - name: charsPerSecond
        dtype: float16
      - name: recording_year-month
        dtype: string
      - name: microphone
        dtype: string
      - name: speaker
        dtype: string
      - name: language
        dtype: string
      - name: comment
        dtype: string
    splits:
      - name: train
        num_bytes: 689946875.552
        num_examples: 2106
    download_size: 627569003
    dataset_size: 689946875.552
configs:
  - config_name: TV-2021.02-Neutral
    data_files:
      - split: train
        path: TV-2021.02-Neutral/train-*
  - config_name: TV-2021.06-Emotional
    data_files:
      - split: train
        path: TV-2021.06-Emotional/train-*
  - config_name: TV-2022.10-Neutral
    data_files:
      - split: train
        path: TV-2022.10-Neutral/train-*
  - config_name: TV-2023.09-Hessisch
    data_files:
      - split: train
        path: TV-2023.09-Hessisch/train-*
  - config_name: all
    description: Meta config or subset containing all four Thorsten-Voice datasets
    homepage: https://www.Thorsten-Voice.de
    license: CC0
    data_files:
      - split: train
        path:
          - TV-2021.02-Neutral/train-*
          - TV-2023.09-Hessisch/train-*
          - TV-2022.10-Neutral/train-*
          - TV-2021.06-Emotional/train-*
license: cc0-1.0
task_categories:
  - text-to-speech
  - text-to-audio
language:
  - de
size_categories:
  - 10K<n<100K

The "Thorsten-Voice" dataset

This truly open source (CC0 license) german (🇩🇪) voice dataset contains about 40 hours of transcribed voice recordings by Thorsten Müller, a single male, native speaker in over 38.000 wave files.

• Mono
• Samplerate: 44.100Hz
• Trimmed silence at begin/end
• Denoised
• Normalized to -24dB

Disclaimer

"Please keep in mind, I am not a professional speaker, just an open source speech technology enthusiast who donates his voice. I contribute my personal voice as a person believing in a world where all people are equal. No matter of gender, sexual orientation, religion, skin color and geocoordinates of birth location. A global world where everybody is warmly welcome on any place on this planet and open and free knowledge and education is available to everyone." (Thorsten Müller)

Features (attributes)

This dataset contains following attributes.

• audio: Actual audio file content. Is playable directly in the browser.
• id: Unique identifier for each file. Format is "RecordingSessionGUID---WaveGUID". Adding a ".wav" to id will be the filename of recording.
• subset: Name of recording session (details below).
• style: Can be "neutral", "hessisch" (a german dialect) or a specific emotion (details below).
• text: The recorded text in this wave file.
• samplerate: The samplerate (44.100Hz) for the recording.
• durationSeconds: Duration for audio file in seconds (rounded to 2).
• charsPerSecond: The recording speed in characters spoken per second.
• recording_year-month: In which month has text been recorded.
• microphone: Some recordings has been made with a bad USB headset and some with a good Rode Podcaster microphone.
• speaker: Guude 👋, it's me - Thorsten 😊.
• language: All recordings are done in german language.
• comment: Some (emotional) recordings might have cut off endings. This is written as comment on affected files.

Subsets & styles

Subset: TV-2021.02-Neutral

This subset contains about 22.000 recordings in a neutral style. The recording quality is mixed. Bad USB microphone or good Rode Podcaster microphone and used a recording chamber. See microphone feature for information. The pronounciation is very clear and slow. Every word is pronounced very well, but the speech flow is less natural because of very clear recording. This subset (in 22kHz samplerate) is also available on Zenodo under DOI 10.5281/zenodo.5525342

Subset: TV-2022.10-Neutral

This subset contains about 12.000 recordings in a neutral style. All recordings where done using a good Rode Podcaster microphone and a recording chamber. The speech flow is very natural. This subset (in 22kHz samplerate) is also available on Zenodo under DOI 10.5281/zenodo.7265581

Subset: TV-2021.06-Emotional

This subset contains about 2.000 recordings in an emotional style. The recorded phrases are for all emotions identical but are pronounced in following different emotions. Some recordings might be cut off too early. This subset (in 22kHz samplerate) is also available on Zenodo under DOI 10.5281/zenodo.5525023

• neutral
• surprised (style: surprised | überrascht)
• disgusted (style: disgusted | angewidert)
• drunk, taken sober (style: drunk | angetrunken)
• angry (style: angry | wütend)
• amused (style: amused | amüsiert)
• whisper (style: whisper | flüstern)
• sleepy (style: sleepy | schläfrig)

Subset: TV-2023.09-Hessisch

This subset contains about 2.000 recordings in a Hessisch (Guude aka. "Hi" 👋). Hessisch is a regional dialect spoken in the state of "Hessen" in the center region of germany. All recordings where done using a good Rode Podcaster microphone and a recording chamber. The speech flow is very natural. This subset (in 22kHz samplerate) is also available on Zenodo under DOI 10.5281/zenodo.10511260

Use the dataset

API Call

You can query the dataset using HuggingFace API with SQL query syntax

SELECT * FROM tv_202106_emotional WHERE "style" = 'angry | wütend' LIMIT 10;

Python Code

from datasets import load_dataset
from datasets import load_dataset_builder
from datasets import get_dataset_config_names

# Get a list of available configs/subsets of Thorsten-Voice dataset
configs_subsets = get_dataset_config_names("Thorsten-Voice/TV-44kHz-Full")
print(configs_subsets)
>>> ['TV-2021.02-Neutral', 'TV-2021.06-Emotional', 'TV-2022.10-Neutral', 'TV-2023.09-Hessisch', 'all']

# Get some dataset information
ds_builder = load_dataset_builder("Thorsten-Voice/TV-44kHz-Full", "TV-2022.10-Neutral")
print("Desciption: " + ds_builder.info.description)
print("Homepage: " + ds_builder.info.homepage)
print("License: " + ds_builder.info.license)
>>> Desciption: Single german male speaker, neutral speech, very clear, high class quality, natural speech flow
>>> Homepage: https://www.Thorsten-Voice.de
>>> License: CC0

# Load "Hessisch" subset
ds = load_dataset("Thorsten-Voice/TV-44kHz-Full", "TV-2023.09-Hessisch", split="train")

# Return first row of "Hessisch" subset
print(ds[0])

# Get first three rows, limited to "text" column
print(ds[:3]["text"])
>>> ['Woran kannst du erkennen, ob etwas qualitativ gut oder schlecht ist.', 'Diese heiße Schokolade ist nichts für Kinder und perfekt, um am Wochenende oder nach einem langen Tag zu entspannen.', 'Aus den Untersuchungen kam heraus, dass diese Kinder aufmerksamer waren, emotional stabiler und ausgeglichener im Vergleich zu den Kindern die später ins Bett gingen.']

Verify dataset integrity

• https://datasets-server.huggingface.co/is-valid?dataset=Thorsten-Voice/TV-44kHz-Full
• https://datasets-server.huggingface.co/statistics?dataset=Thorsten-Voice/TV-44kHz-Full&config=TV-2021.02-Neutral&split=train

DOI / cite

@misc {thorsten_müller_2024,
    author       = { {Thorsten Müller} },
    title        = { TV-44kHz-Full (Revision ff427ec) },
    year         = 2024,
    url          = { https://huggingface.co/datasets/Thorsten-Voice/TV-44kHz-Full },
    doi          = { 10.57967/hf/3290 },
    publisher    = { Hugging Face }
}

Thanks

Thanks to all amazing open source communities around the globe for bringing the world forward. Of course, thanks to Dominik Kreutz for your great support over the years 🤗.

Links

• https://www.Thorsten-Voice.de
• https://www.youtube.com/@ThorstenMueller
• https://github.com/thorstenMueller/Thorsten-Voice
• https://huggingface.co/spaces/Thorsten-Voice/TTS