Skit-S2I Dataset

Dataset Release for Intent Classification task from Speech

About

This is a dataset for Intent classification from human speech, and covers 14 coarse-grained intents from the Banking domain. This work is inspired by a similar release in the Minds-14 dataset - here, we restrict ourselves to Indian English but with a much larger training set. The dataset is split into:

• test - 100 samples per intent
• train - >650 samples per intent

The data was generated by 11 (Indian English) speakers, recording over a telephony line. We also provide access to anonymised speaker information - like gender, languages spoken, native language - so as to allow more structured discussions around robustness and bias, in the models you train.

Download and Usage

The dataset is available on HuggingFace as Skit-S2I.

This dataset is shared under Creative Commons Attribution-NonCommercial 4.0 International Licence. This places restrictions on commercial use of this dataset.

Uses

Most spoken dialog-systems use a pipeline of speech recognition followed by intent classification, and optimise each individually. But this allows ASR errors to leak downstream. Instead, what if we train end-to-end intent models on speech ? More importantly, how well would such models generalise in a language like Indian English - given the diversity of speech behaviours ? This dataset is an attempt towards answering such questions around robustness and model bias.

Structure

This release contains data of (Indian English) speech samples tagged with an intent from the Banking domain. Also includes the transcript template used to generate the sample.

Audio Quality : 8 Khz, 16-bit

Structure

- wav_audios          [contains the wav audio files]
- train.csv           [contains the train split, where each row contains "<id> | <intent-class> | <template> | <audio-path> | <speaker-id>"]
- test.csv            [contains the test split, where each row contains "<id> | <intent-class> | <template> | <audio-path> | <speaker-id>"]
- intent_info.csv     [contains information about the intents, where each row contains "<intent-class> | <intent-name> | <description>"]
- speaker_info.csv    [contains information about the speakers, where each row contains "<speaker-id> | <native-language> | <languages-spoken> | <places-lived> | <gender>"]

More information regarding the dataset can be found in the datasheet.

Baselines

The code for the baselines are provided in the baselines directory.

Citation

If you are using this dataset, please cite using the link in the About section on the right.

License

Shield:

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.