metadata

dataset_info:
  features:
    - name: audio
      dtype: audio
    - name: transcription
      dtype: string
  splits:
    - name: train
      num_bytes: 74359398844.398
      num_examples: 292462
  download_size: 107587361654
  dataset_size: 74359398844.398
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
task_categories:
  - automatic-speech-recognition
language:
  - de

Dataset Card for "bundestagv2"

Table of Contents
Dataset Description
- Dataset Summary
- Supported Tasks and Leaderboards
Dataset Structure
Dataset Creation
- Source Data
Considerations for Using the Data
- Discussion of Biases
- Other Known Limitations
Additional Information
- Licensing Information

Dataset Description

This is the second version of our bundestag Dataset and the first one that reached a sufficient quality. It has been obtained using the .srt files found on the website of "Parlamentsfernsehen Bundestag" and has been further optimized using forced alignment.

Dataset Summary

Almost 300k Rows of data, filtered to up to 30 seconds of audio.

Supported Tasks and Leaderboards

This dataset is mostly intended for ASR usage.

Dataset Structure

Data Instances

Every instance ranges from 0 to 30 seconds of audio.

Data Fields

We provide the waveforms and the corresponding transcriptions.

Data Splits

This is currently one large split, so consider applying train_test_split first before using it for training.

Dataset Creation

Source Data

Initial Data Collection and Normalization

Aligned using oliverguhr/wav2vec2-large-xlsr-53-german-cv9

Considerations for Using the Data

Discussion of Biases

There are quite a few dialects present throughout the dataset.

Other Known Limitations

A lot of the sessions were held online, as such the audio quality varies drastically and is sometimes too bad to ensure good transcriptions. The .srt files used are not flawless, as such there might be a few errors in the dataset. Additionally, the alignment will be off from time to time, resulting in missing speech with the full transcript which could increase hallucinations.

Additional Information

Licensing Information

This dataset is currently gated and will remain this as long as we are uncertain about the legality of republishing audio of the German "Bundestag".