Mathematical Sciences

Lund University

  • Title: How many persons are there in the room?
  • Short description:

    Minut AB makes a smart home alarm and monitoring system that is increasingly used as a way for short-stay hosts to monitor their properties. Our system has been designed from the ground up with privacy in mind, which makes it a good fit for people who want to make sure their properties are all right, without having a camera there. The reasons for why short-stay hosts might want to monitor their properties are many, one reason is that it's increasingly becoming a problem that guests are hosting parties in the rented properties, sometimes with devastating outcomes (https://www.bbc.com/news/world-us-canada-50276485).

    There are many ways in which a device like Minut could figure out whether there is a party ongoing in a property, the most telling signal would probably be the level of noise. Another possibility is to analyse the frequency content of the audio to get an understanding of what is happening, this is something we at Minut have done before by classifying the 'acoustic scene' as either a party or not.

    One factor that could determine whether there is a risk for a party is the number of people in the property. The number of people in the property is also useful for the host to know even if there is no party ongoing, since the billing is often based on the number of people that are staying over.

    We would like to explore the possiblity to use audio analysis to get an estimate of the number of people that are nearby, by counting the number of unique voices heard in an acoustic scene. We think the best approach for doing this is by training deep neural networks on some representation of the audio, but the student(s) are free to use which methods they want. One twist to this problem is that we would like to run the trained model on a small microcontroller, which puts serious contraints on the size and runtime memory usage by the model.

    The candidate(s) should preferably have a strong interest in machine learning and digital signal processing. We would be happy to see two students working on this, but you are welcome to apply alone too. If you are interested, send an email to colin@minut.com and tell us about your interests.

  • Long description:

    Minut AB makes a smart home alarm and monitoring system that is increasingly used as a way for short-stay hosts to monitor their properties. Our system has been designed from the ground up with privacy in mind, which makes it a good fit for people who want to make sure their properties are all right, without having a camera there. The reasons for why short-stay hosts might want to monitor their properties are many, one reason is that it's increasingly becoming a problem that guests are hosting parties in the rented properties, sometimes with devastating outcomes (https://www.bbc.com/news/world-us-canada-50276485).

    There are many ways in which a device like Minut could figure out whether there is a party ongoing in a property, the most telling signal would probably be the level of noise. Another possibility is to analyse the frequency content of the audio to get an understanding of what is happening, this is something we at Minut have done before by classifying the 'acoustic scene' as either a party or not.

    One factor that could determine whether there is a risk for a party is the number of people in the property. The number of people in the property is also useful for the host to know even if there is no party ongoing, since the billing is often based on the number of people that are staying over.

    We would like to explore the possiblity to use audio analysis to get an estimate of the number of people that are nearby, by counting the number of unique voices heard in an acoustic scene. We think the best approach for doing this is by training deep neural networks on some representation of the audio, but the student(s) are free to use which methods they want. One twist to this problem is that we would like to run the trained model on a small microcontroller, which puts serious contraints on the size and runtime memory usage by the model.

    The candidate(s) should preferably have a strong interest in machine learning and digital signal processing. We would be happy to see two students working on this, but you are welcome to apply alone too. If you are interested, send an email to colin@minut.com and tell us about your interests.

  • Industrial cooperation: Minut AB
  • Contact: Andreas Jakobsson