Replication Data for: Self-Learning for Personalized Keyword Spotting on Ultra-Low-Power Audio Sensors

This repository includes the data for reproducing the results of the paper: "Self-Learning for Personalized Keyword Spotting on Ultra-Low-Power Audio Sensors". In this work, we recorded a dataset of speech samples with our microphone sensor after replaying a subset of data from the HeySnips dataset using a speaker.
The collected speech data (total of 400 samples) are split between a testset and a trainset, both including "Hey Snips" utterances and non-"Hey Snips" utterances.
In particular, the data of the testset is composed by recordings from 20 random speakers from the original testset.
After the recording, the data were fed to our DNN models deployed on devices.
Initially, a per-speaker prototype vector is computed by feeding three audio recordings of the target keywords.
Next, the audio tracks of the training set are processed with a sliding window approach to compute the distance with respect to the prototype and assign pseudo-labels for the self-learning task.

The dataset is, therefore, composed of two main partitions. First, the "recorded_speech_data" includes the audio recordings. Note that this dataset is under restricted access to not violate the terms of access of the original dataset. Second, the "processed_outputs" includes the output of the processing, i.e. the measured distances.
By using the dataset in combination with the associated code, every user will be able to reproduce the results of the paper.