Dataset Card for SLUE

Dataset Summary

We introduce the Spoken Language Understanding Evaluation (SLUE) benchmark. The goals of our work are to

• Track research progress on multiple SLU tasks
• Facilitate the development of pre-trained representations by providing fine-tuning and eval sets for a variety of SLU tasks
• Foster the open exchange of research by focusing on freely available datasets that all academic and industrial groups can easily use.

For this benchmark, we provide new annotation of publicly available, natural speech data for training and evaluation. We also provide a benchmark suite including code to download and pre-process the SLUE datasets, train the baseline models, and evaluate performance on SLUE tasks. Refer to Toolkit and Paper for more details.

Supported Tasks and Leaderboards

Automatic Speech Recognition (ASR)

Although this is not a SLU task, ASR can help analyze the performance of downstream SLU tasks on the same domain. Additionally, pipeline approaches depend on ASR outputs, making ASR relevant to SLU. ASR is evaluated using word error rate (WER).

Named Entity Recognition (NER)

Named entity recognition involves detecting the named entities and their tags (types) in a given sentence. We evaluate performance using micro-averaged F1 and label-F1 scores. The F1 score evaluates an unordered list of named entity phrase and tag pairs predicted for each sentence. Only the tag predictions are considered for label-F1.

Sentiment Analysis (SA)

Sentiment analysis refers to classifying a given speech segment as having negative, neutral, or positive sentiment. We evaluate SA using macro-averaged (unweighted) recall and F1 scores.[More Information Needed]

How-to-submit for your test set evaluation

See here https://asappresearch.github.io/slue-toolkit/how-to-submit.html

Languages

The language data in SLUE is in English.

Dataset Structure

Data Instances

voxpopuli

• Size of downloaded dataset files: 398.45 MB
• Size of the generated dataset: 5.81 MB
• Total amount of disk used: 404.26 MB An example of 'train' looks as follows.

{'id': '20131007-0900-PLENARY-19-en_20131007-21:26:04_3',
 'audio': {'path': '/Users/username/.cache/huggingface/datasets/downloads/extracted/e35757b0971ac7ff5e2fcdc301bba0364857044be55481656e2ade6f7e1fd372/slue-voxpopuli/fine-tune/20131007-0900-PLENARY-19-en_20131007-21:26:04_3.ogg',
  'array': array([ 0.00132601,  0.00058881, -0.00052187, ...,  0.06857217,
          0.07835515,  0.07845446], dtype=float32),
  'sampling_rate': 16000},
 'speaker_id': 'None',
 'normalized_text': 'two thousand and twelve for instance the new brussels i regulation provides for the right for employees to sue several employers together and the right for employees to have access to courts in europe even if the employer is domiciled outside europe. the commission will',
 'raw_text': '2012. For instance, the new Brussels I Regulation provides for the right for employees to sue several employers together and the right for employees to have access to courts in Europe, even if the employer is domiciled outside Europe. The Commission will',
 'raw_ner': {'type': ['LOC', 'LOC', 'LAW', 'DATE'],
  'start': [227, 177, 28, 0],
  'length': [6, 6, 21, 4]},
 'normalized_ner': {'type': ['LOC', 'LOC', 'LAW', 'DATE'],
  'start': [243, 194, 45, 0],
  'length': [6, 6, 21, 23]},
 'raw_combined_ner': {'type': ['PLACE', 'PLACE', 'LAW', 'WHEN'],
  'start': [227, 177, 28, 0],
  'length': [6, 6, 21, 4]},
 'normalized_combined_ner': {'type': ['PLACE', 'PLACE', 'LAW', 'WHEN'],
  'start': [243, 194, 45, 0],
  'length': [6, 6, 21, 23]}}

voxceleb

• Size of downloaded dataset files: 1.55 GB
• Size of the generated dataset: 3.78 MB
• Total amount of disk used: 1.55 GB An example of 'train' looks as follows.

{'id': 'id10059_229vKIGbxrI_00004',
 'audio': {'path': '/Users/felixwu/.cache/huggingface/datasets/downloads/extracted/400facb6d2f2496ebcd58a5ffe5fbf2798f363d1b719b888d28a29b872751626/slue-voxceleb/fine-tune_raw/id10059_229vKIGbxrI_00004.flac',
  'array': array([-0.00442505, -0.00204468,  0.00628662, ...,  0.00158691,
          0.00100708,  0.00033569], dtype=float32),
  'sampling_rate': 16000},
 'speaker_id': 'id10059',
 'normalized_text': 'of god what is a creator the almighty that uh',
 'sentiment': 'Neutral',
 'start_second': 0.45,
 'end_second': 4.52}

Data Fields

voxpopuli

• id : a string id of an instance.
• audio : audio feature of the raw audio. It is a dictionary containing the path to the downloaded audio file, the decoded audio array, and the sampling rate. Note that when accessing the audio column: dataset[0]["audio"] the audio file is automatically decoded and resampled to dataset.features["audio"].sampling_rate . Decoding and resampling of a large number of audio files might take a significant amount of time. Thus it is important to first query the sample index before the "audio" column, i.e. dataset[0]["audio"] should always be preferred over dataset["audio"][0] .
• speaker_id : a string of the speaker id.
• raw_text : a string feature that contains the raw transcription of the audio.
• normalized_text : a string feature that contains the normalized transcription of the audio which is used in the standard evaluation .
• raw_ner : the NER annotation of the raw_text using the same 18 NER classes as OntoNotes.
• normalized_ner : the NER annotation of the normalized_text using the same 18 NER classes as OntoNotes.
• raw_combined_ner : the NER annotation of the raw_text using our 7 NER classes ( WHEN , QUANT , PLACE , NORP , ORG , LAW , PERSON ).
• normalized_combined_ner : the NER annotation of the normalized_text using our 7 NER classes ( WHEN , QUANT , PLACE , NORP , ORG , LAW , PERSON ) which is used in the standard evaluation . Each NER annotation is a dictionary containing three lists: type , start , and length . type is a list of the NER tag types. start is a list of the start character position of each named entity in the corresponding text. length is a list of the number of characters of each named entity.

voxceleb

• id : a string id of an instance.
• audio : audio feature of the raw audio. Please use start_second and end_second to crop the transcribed segment. For example, dataset[0]["audio"]["array"][int(dataset[0]["start_second"] * dataset[0]["audio"]["sample_rate"]):int(dataset[0]["end_second"] * dataset[0]["audio"]["sample_rate"])] . It is a dictionary containing the path to the downloaded audio file, the decoded audio array, and the sampling rate. Note that when accessing the audio column: dataset[0]["audio"] the audio file is automatically decoded and resampled to dataset.features["audio"].sampling_rate . Decoding and resampling of a large number of audio files might take a significant amount of time. Thus it is important to first query the sample index before the "audio" column, i.e. dataset[0]["audio"] should always be preferred over dataset["audio"][0] .
• speaker_id : a string of the speaker id.
• normalized_text : a string feature that contains the transcription of the audio segment.
• sentiment : a string feature which can be Negative , Neutral , or Positive .
• start_second : a float feature that specifies the start second of the audio segment.
• end_second : a float feature that specifies the end second of the audio segment.

Data Splits

train	validation	test
voxpopuli	5000	1753	1842
voxceleb	5777	1454	3553
Here we use the standard split names in Huggingface's datasets, so the train and validation splits are the original fine-tune and dev splits of SLUE datasets, respectively.

Dataset Creation

Curation Rationale

[More Information Needed]

Source Data

Initial Data Collection and Normalization

[More Information Needed]

Who are the source language producers?

[More Information Needed]

Annotations

Annotation process

[More Information Needed]

Who are the annotators?

[More Information Needed]

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

SLUE-VoxPopuli Dataset

SLUE-VoxPopuli dataset contains a subset of VoxPopuli dataset and the copyright of this subset remains the same with the original license, CC0. See also European Parliament's legal notice ( https://www.europarl.europa.eu/legal-notice/en/ )

Additionally, we provide named entity annotation (normalized_ner and raw_ner column in .tsv files) and it is covered with the same license as CC0.

SLUE-VoxCeleb Dataset

SLUE-VoxCeleb Dataset contains a subset of OXFORD VoxCeleb dataset and the copyright of this subset remains the same Creative Commons Attribution 4.0 International license as below. Additionally, we provide transcription, sentiment annotation and timestamp (start, end) that follows the same license to OXFORD VoxCeleb dataset.

Original License of OXFORD VGG VoxCeleb Dataset

VoxCeleb1 contains over 100,000 utterances for 1,251 celebrities, extracted from videos uploaded to YouTube. VoxCeleb2 contains over a million utterances for 6,112 celebrities, extracted from videos uploaded to YouTube.

The speakers span a wide range of different ethnicities, accents, professions and ages.

We provide Youtube URLs, associated face detections, and timestamps, as well as cropped audio segments and cropped face videos from the dataset. The copyright of both the original and cropped versions of the videos remains with the original owners.

The data is covered under a Creative Commons Attribution 4.0 International license (Please read the license terms here. https://creativecommons.org/licenses/by/4.0/ ).

Downloading this dataset implies agreement to follow the same conditions for any modification and/or re-distribution of the dataset in any form.

Additionally any entity using this dataset agrees to the following conditions:

THIS DATASET IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

Please cite [1,2] below if you make use of the dataset.

[1] J. S. Chung, A. Nagrani, A. Zisserman VoxCeleb2: Deep Speaker Recognition INTERSPEECH, 2018.

[2] A. Nagrani, J. S. Chung, A. Zisserman VoxCeleb: a large-scale speaker identification dataset INTERSPEECH, 2017

Citation Information

@inproceedings{shon2022slue,
  title={Slue: New benchmark tasks for spoken language understanding evaluation on natural speech},
  author={Shon, Suwon and Pasad, Ankita and Wu, Felix and Brusco, Pablo and Artzi, Yoav and Livescu, Karen and Han, Kyu J},
  booktitle={ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={7927--7931},
  year={2022},
  organization={IEEE}
}

Contributions

Thanks to @fwu-asapp for adding this dataset.