We introduce two possible Machine Learning approaches to address the problem of intent detection. The solution compares different classification models: Random Forests and Support Vector Machines. By leveraging both time and frequency-based features of the audio files, both models achieve promising results.
Intent detection aims to recognize the action that the user wants to perform. The proposed task focuses on classifying the user’s intent, given a dataset of spoken utterances. The dataset consists of two parts:
- Audio files: audio samples in .wav format
- Metadata: files .csv that contain information about each audio sample in the dataset and about its speaker.
The dataset provided is composed of 11,309 recordings (.wav) of user queries with English pronunciation. The metadata have been distributed in two separate collections:
- a development set, containing 9,854 records of recordings for which the intent is known;
- an evaluation set, containing 1,455 records of recordings without the target variable.
We used the development set to build a classification model capable of classifying user intentions for each record of the evaluation set.