metadata

license: apache-2.0
dataset_info:
  features:
    - name: hypothesis
      sequence: string
    - name: transcription
      dtype: string
    - name: input1
      dtype: string
    - name: hypothesis_concatenated
      dtype: string
    - name: source
      dtype: string
    - name: id
      dtype: string
    - name: dummy_str
      dtype: string
    - name: dummy_list
      sequence: 'null'
    - name: prompt
      dtype: string
  splits:
    - name: train
      num_bytes: 469086507
      num_examples: 286366
    - name: test
      num_bytes: 24103011
      num_examples: 18237
  download_size: 125101353
  dataset_size: 493189518
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*

Dataset Name: Pilot dataset for Multi-domain ASR corrections

Description

This dataset is a pilot version of a larger dataset for automatic speech recognition (ASR) corrections across multiple domains. It contains paired hypotheses and corrected transcriptions for various ASR tasks consolidated from PeacefulData/HyPoradise-v0

Structure

Data Split

The dataset is divided into training and test splits:

Training Data: 281,082 entries
- Approximately 6,255,198 tokens for transcriptions
- Approximately 31,211,083 tokens for concatenated hypotheses
Test Data: 16,108 entries
- Approximately 327,750 tokens for transcriptions
- Approximately 1,629,093 tokens for concatenated hypotheses

Columns

hypothesis: N-best hypothesis from beam search.
transcription: Corrected asr transcription.
hypothesis_concatenated: An alternative version of the text output.
source: The source of the text entry, indicating the origin dataset.
prompt: Instructional prompt for correction task
score: An acoustic model score (not all entries have this).

Source Datasets

The dataset combines entries from various sources:

Training Sources:
- train_td3: 50,000 entries
- train_other_500: 50,000 entries
- train_cv: 47,293 entries
- train_lrs2: 42,940 entries
- train_wsj_score: 37,514 entries
- train_swbd: 36,539 entries
- train_chime4: 9,600 entries
- train_atis: 3,964 entries
- train_coraal: 3,232 entries
Test Sources:
- test_ls_other: 2,939 entries
- test_ls_clean: 2,620 entries
- test_lrs2: 2,259 entries
- test_swbd: 2,000 entries
- test_cv: 2,000 entries
- test_chime4: 1,320 entries
- test_td3: 1,155 entries
- test_wsj_score: 836 entries
- test_atis: 809 entries
- test_coraal: 170 entries

Access

The dataset can be accessed and downloaded through the HuggingFace Datasets library. Use the following command to load the dataset:

from datasets import load_dataset
dataset = load_dataset("PeacefulData/HyPoradise-pilot")

Acknowledgments

This dataset is consolidated from the PeacefulData/HyPoradise-v0 dataset. Thanks to the original creators for making this data available.

References

@inproceedings{yang2023generative,
  title={Generative speech recognition error correction with large language models and task-activating prompting},
  author={Yang, Chao-Han Huck and Gu, Yile and Liu, Yi-Chieh and Ghosh, Shalini and Bulyko, Ivan and Stolcke, Andreas},
  booktitle={2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)},
  pages={1--8},
  year={2023},
  organization={IEEE}
}

@inproceedings{chen2023hyporadise,
  title={HyPoradise: An Open Baseline for Generative Speech Recognition with Large Language Models},
  author={CHEN, CHEN and Hu, Yuchen and Yang, Chao-Han Huck and Siniscalchi, Sabato Marco and Chen, Pin-Yu and Chng, Ensiong},
  booktitle={Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
  year={2023}
}