The viewer is disabled because this dataset repo requires arbitrary Python code execution. Please consider removing the loading script and relying on automated data support (you can use convert_to_parquet from the datasets library). If this is not possible, please open a discussion for direct help.

Dataset Card for ChaLL

Dataset Summary

This dataset contains audio recordings of spontaneous speech by young learners of English in Switzerland. The recordings capture various language learning tasks designed to elicit authentic communication from the students. The dataset includes detailed verbatim transcriptions with annotations for errors made by the learners. The transcripts were prepared by a professional transcription service, and each recording was associated with detailed metadata, including school grade, recording conditions, and error annotations.

Data Availability: The dataset that we collected contains sensitive data of minors and thus cannot be shared publicly. The data can, however, be accessed as part of a joint project with one or several of the original project partners, subject to a collaboration agreement (yet to be detailed).

To use the ChaLL dataset, you need to download it manually. Once you have manually downloaded the files, please extract all files into a single folder. You can then load the dataset into your environment using the following command:

from datasets import load_dataset
dataset = load_dataset('chall', data_dir='path/to/folder/folder_name')

Ensure the path specified in data_dir correctly points to the folder where you have extracted the dataset files.

Examples in this dataset are generated using the soundfile library (for reading and chunking). To handle the audio data correctly, you need to install the soundfile library in your project.

pip install soundfile

Supported Tasks and Leaderboards

[More Information Needed]

Languages

The primary language represented in this dataset is English, specifically as spoken by Swiss children who are learners of the language. This includes a variety of accents and dialectal influences from the German-speaking regions of Switzerland.

Dataset Structure

The dataset can be loaded using different configurations to suit various experimental setups.

Dataset Builder Configuration

The configurations define how the data is preprocessed and loaded into the environment. Below are the details of the configurations used in experiments:

original

This configuration uses the data in its raw, unmodified form while ensuring all participant information is anonymized. It includes the preservation of the data's original structure without segmentation, filtering, or other preprocessing techniques.

from datasets import load_dataset
dataset = load_dataset('mict-zhaw/chall', 'original', data_dir='path/to/folder/folder_name')

asr

This configuration is intended for ASR experiments, enabling segment splitting for more granular processing of the audio data.

asr_acl

This configuration includes specific settings used in the related research paper. It is designed to handle various segmentation and preprocessing tasks to prepare the data.

The results for the paper were generated at a time when the data was not yet complete. Thus, this dataset configuration comprises approximately 85 hours (excluding pauses between utterances) of spontaneous English speech recordings from young learners in Switzerland, collected from 327 distinct speakers in grades 4 to 6. The dataset includes 45,004 individual utterances and is intended to train an ASR system that preserves learner errors for corrective feedback in language learning.

The configuration is set to split segments, with a maximum pause length of 12 seconds, maximum chunk length of 12 seconds, minimum chunk length of 0.5 seconds, removes trailing pauses, converts text to lowercase and numbers to words.

from datasets import load_dataset
dataset = load_dataset('mict-zhaw/chall', 'asr_acl', data_dir='path/to/folder/folder_name')

Custom

The ChallConfig class provides various parameters that can be customized:

  • split_segments (bool): Whether to split the audio into smaller segments.
  • max_chunk_length (float or None): Maximum length of each audio chunk in seconds (used only if split_segments is True).
  • min_chunk_length (float or None): Minimum length of each audio chunk in seconds (used only if split_segments is True).
  • max_pause_length (float or None): Maximum allowable pause length within segments (used only if split_segments is True).
  • remove_trailing_pauses (bool): Whether to remove trailing pauses from segments (used only if split_segments is True).
  • lowercase (bool): Whether to convert all text to lowercase.
  • num_to_words (bool): Whether to convert numerical expressions to words.
  • allowed_chars (set): Set of allowed characters in the text. Automatically set based on the lowercase parameter.
  • special_terms_mapping (dict): Dictionary for mapping special terms to their replacements.
  • stratify_column (str or None): Column used for stratifying the data into different folds.
  • folds (dict or None): Dictionary defining the data folds for stratified sampling.

Custom configurations can be used alone or in combination with existing ones, and they will overwrite predefined defaults.

from datasets import load_dataset
dataset = load_dataset('mict-zhaw/chall', data_dir='path/to/folder/folder_name', **kwargs)
dataset = load_dataset('mict-zhaw/chall', 'asr_acl', data_dir='path/to/folder/folder_name', **kwargs)

Data Instances

A typical data instance in this dataset include an audio file, its full transcription, error annotations, and associated metadata such as the speaker's grade level and recording conditions. Here is an example:

split_segments == True

When split_segments is set to True, the audio data is divided into utterances. An utterance data instance includes the spoken text from one participant along with meta information such as school_grade, area_of_school_code, background_noise, and intervention. The audio is present as byte array under audio.

{
  "audio_id": "S004_A005_000", 
  "intervention": 4, 
  "school_grade": "6", 
  "area_of_school_code": 5, 
  "background_noise": false, 
  "raw_text": "A male or is it a female?", 
  "clear_text": "a male or is it a female", 
  "words": {
    "start": [0.4099999964237213, 0.5600000023841858, 1.0399999618530273, 1.25, 1.3700000047683716, 1.5499999523162842, 1.6699999570846558],
    "end": [0.5400000214576721, 1.0399999618530273, 1.25, 1.3700000047683716, 1.5499999523162842, 1.6699999570846558, 2.5399999618530273],
    "duration": [0.1300000250339508, 0.47999995946884155, 0.21000003814697266, 0.12000000476837158, 0.1799999475479126, 0.12000000476837158, 0.8700000047683716],
    "text": ["A", "male", "or", "is", "it", "a", "female?"]
  }, 
  "audio": {
    "path": false, 
    "array": [0, 0, 0, "...", 0, 0, 0], 
    "sampling_rate": 16000
  }
}

split_segments == False

When split_segments is set to False, the audio remains intact and includes multiple turns with one or more speakers. In this case, additional participant meta information is present, but speakers (from the transcript) and participants cannot be aligned and do not need to match in number. This means the transcription agency may define more than one speaker for a single participant.

{
  "audio_id": "S001_A046",
  "intervention": 1,
  "school_grade": "4",
  "area_of_school_code": 2,
  "raw_text": "If you could have-have any superpower, what would it be? I would choose to have invincibility because when I'm invincible, I can't die or get hurt by anyone and I think this concept is very cool...",
  "clear_text": "if you could have have any superpower what would it be i would choose to have invincibility because when i'm invincible i can't die or get hurt by anyone and i think this concept is very cool...", 
  "participants": {
    "estimated_l2_proficiency": [null, null], 
    "gender": ["M", "F"], 
    "languages": ["NNS", "NNS"], 
    "pseudonym": ["P033", "P034"], 
    "school_grade": [6, 6], 
    "year_of_birth": [2010, 2011]
  }, 
  "background_noise": true, 
  "speakers": {
    "name": ["Participant 1", "Participant 2"], 
    "spkid": ["S002_A004_SPK0", "S002_A004_SPK1"]
  }, 
  "segments": {
      "speaker": ["S002_A004_SPK0", "S002_A004_SPK1", ...],
      "words": [
        {
          "start": [1.8799999952316284, 2.119999885559082, 3.2899999618530273, ...],
          "end": [2.119999885559082, 2.390000104904175, 3.859999895095825, ...],
          "duration": [0.2399998903274536, 0.2700002193450928, 0.5699999332427979, ...],
          "text": ["If", "you", "could", "have-have", "any", "superpower,", "what", "would", "it", "be?"]
        },       {
          "start": [10.760000228881836, 11.029999732971191, 11.420000076293945, ...],
          "end": [11.029999732971191, 11.420000076293945, 12.170000076293945, ...],
          "duration": [0.26999950408935547, 0.3900003433227539, 0.75, ...],
          "text": ["I", "would", "choose", "to", "have", "invincibility", "because", "when", "I'm", "invincible,", "I", "can't", "die", "or", "get", "hurt", "by", "anyone", "and", "I", "think", "this", "concept", "is", "very", "cool."]
        }
    ]
  }, 
  "audio": {
    "path": null,
    "array": [0, 0, 0, ..., 0, 0, 0], 
    "sampling_rate": 16000
  }
}

Data Fields

  • audio_id: A unique identifier for the audio recording.
  • intervention: An integer representing the type or stage of intervention.
  • school_grade: The grade level of the student(s) involved in the recording.
  • area_of_school_code: A code representing a specific area within the school.
  • raw_text: The raw transcription of the audio, capturing exactly what was spoken.
  • clear_text: A cleaned version of the raw text, formatted for easier analysis.
  • background_noise: A boolean indicating whether background noise is present in the recording.
  • audio: An object containing the audio data and related information.
    • path: The file path of the audio recording (can be null).
    • array: An array representing the audio waveform data.
    • sampling_rate: The rate at which the audio was sampled, in Hz.

In addition to the common fields, there are specific fields depending on split_segments:

Utterance Data Instance (True)

  • words: An object containing details about each word spoken in the utterance.
    • start: A list of start times for each word.
    • end: A list of end times for each word.
    • duration: A list of durations for each word.
    • text: A list of words spoken.

Audio Data Instance (False)

  • participants: An object containing meta information about the participants in the recording.
    • estimated_l2_proficiency: A list of estimated language proficiency levels.
    • gender: A list of genders of the participants.
    • languages: A list of languages spoken by the participants.
    • pseudonym: A list of pseudonyms assigned to the participants.
    • school_grade: A list of school grades for each participant.
    • year_of_birth: A list of birth years for each participant.
  • speakers: An object containing information about the speakers in the transcript.
    • name: A list of speaker names as identified in the transcript.
    • spkid: A list of speaker IDs.
  • segments: An object containing details about each segment of the recording.
    • speaker: A list of speaker IDs for each segment.
    • words: A list of objects, each containing details about the words spoken in the segment.
      • start: A list of start times for each word in the segment.
      • end: A list of end times for each word in the segment.
      • duration: A list of durations for each word in the segment.
      • text: A list of words spoken in the segment.

Data Splits

The data splits can define as part of the configuration using the folds. Without specifying folds all data is loaded in the train split.

asr_acl

For the experiments in this paper, we split the dataset into five distinct folds of similar duration (about 16h each), where each class (and therefore also each speaker) occurs in only one fold. To simulate the use case of the ASR system being confronted with a new class of learners, each fold contains data from a mix of grades. The following figure visualises the duration and grade distribution of each fold.

Chall Folds

Dataset Creation

Curation Rationale

The dataset was created to address the need for ASR systems that can handle children’s spontaneous speech and preserve their errors to provide effective corrective feedback in language learning environments.

Source Data

Initial Data Collection and Normalization

Audio data was collected from primary school students aged 9 to 14 years, performing language learning tasks in pairs, trios, or individually. The recordings were made at schools and universities, and detailed verbatim transcriptions were created by a transcription agency, following specific guidelines.

Who are the source language producers?

The source data producers include primary school students from German-speaking Switzerland, aged 9 to 14 years, participating in language learning activities.

Annotations

Annotation process

The transcription and annotation process was outsourced to a transcription agency, following detailed guidelines for error annotation, including symbols for grammatical, lexical, and pronunciation errors, as well as German word usage.

Who are the annotators?

The annotators were professionals from a transcription agency, trained according to specific guidelines provided by the project team.

Personal and Sensitive Information

The dataset contains audio recordings of minors. All data was collected with informed consent from legal guardians, and recordings are anonymized to protect the identities of the participants.

Considerations for Using the Data

Social Impact of Dataset

The dataset supports the development of educational tools that could enhance language learning for children, providing an important resource for educational technology.

Discussion of Biases

Given the specific demographic (Swiss German-speaking schoolchildren), the dataset may not generalize well to other forms of English or to speakers from different linguistic or cultural backgrounds.

Other Known Limitations

The outsourcing of transcription and error annotations always poses a risk of yielding erroneous data, since most transcribers are not trained in error annotation.

Additional Information

Dataset Curators

The dataset was curated by researchers at PHZH, UZH and Zhaw, with collaboration from local schools in Switzerland.

Licensing Information

[More Information Needed]

Citation Information

@inproceedings{
  anonymous2024errorpreserving,
  title={Error-preserving Automatic Speech Recognition of Young English Learners' Language},
  author={Janick Michot, Manuela Hürlimann, Jan Deriu, Luzia Sauer, Katsiaryna Mlynchyk, Mark Cieliebak},
  booktitle={The 62nd Annual Meeting of the Association for Computational Linguistics},
  year={2024},
  url={https://openreview.net/forum?id=XPIwvlqIfI}
}

Contributions

Thanks to @mict-zhaw for adding this dataset.

Downloads last month
535

Models trained or fine-tuned on mict-zhaw/chall