The interactions between human and machines are now common in our daily life. The audio data of human communication is a rich source of information, but it is con- sidered privacy-invasive for machines to listen to it. By reducing sampling frequency, it is possible to preserve privacy by making conversation unclear while still being possible to detect if someone is speaking or not. The topic of this paper is to investigate how low sampled frequency audio data hinders the detection of speech. To detect speaking, voice activity detection has been applied, which is a technology in the signal process- ing field that identifies which short segments of audio contain speakings. Two types of state-of-art voice activity detector(VAD) were used for this experiment including a supervised (pyannote) and two unsupervised (rVAD pitch and flatness mode) methods. As a result, the unsupervised methods outperformed the supervised model, where rVAD pitch mode has resulted in the best performance out of all three. More specifically, the unsupervised VAD’s performance became lower as the sample rates decreased while the supervised VAD did not work well at higher sample frequency. rVAD pitch mode at sample rates of 8000Hz or higher was possible to perform at the almost same level as a state-of-art supervised VAD that is trained in a similar data set. Furthermore, it was able to perform as well as a modern unsupervised VAD at 2000Hz or higher sample frequencies. At the sample rate of 1250Hz or lower, any VAD was not able to perform at the same level as a state-of-art VAD. Regarding the privacy aspect, it is observed that human ears detect speaking better than computers, where humans can understand parts or all of the contents of speaking at 2000Hz or higher, which infers that current technology is not enough to detect speech from downsampled privacy-preserving audio. However, there is still a need for further research to verify the effects of the training set and its sample frequencies for the supervised method and also proper scientific so- cial experiments to test the ability of humans of speech detection for reduced sampled audio.

With the development of new technologies and approaches in the field of social signal processing, concerns regarding privacy around recording conversations have arised. One of the main ways to preserve the privacy of the speakers in recorded conversations consists of decimating said conversations, which consists of reducing the sample frequency and the frequency bandwidth of the audio. This theoretically makes the verbal content of the conversation (the words themselves) unintelligible, while still preserving other useful non-verbal social cues such as laughter, pitch modulation and frequency of speech, amongst others. However, this has not been experimentally verified. This research paper addresses this knowledge gap by exploring the performance of laughter detection machine learning models with decimated audio. An existing pre-trained state-of-the-art laughter detection model was employed and its performance was evaluated for a dataset of decimated audio with sample frequencies ranging from 300Hz to 44100Hz.

Living in a world where every single electronic device is online and interconnected, privacy is a growing concern. Finding the threshold where audio is unintelligible to transcription software is crucial when everything that we say can be recorded. Even if Automated Speech Recognition (ASR) is used in tools, such as Siri or Alexa, designed to ease daily tasks, it could also be used in malicious manners. ASR technology has not been around for too long and like any other new piece of technology, it still has many aspects that have not been looked into and are unknown to the public. This research paper addresses this knowledge gap by examining how sample frequency reduction affects word detection using current well-known transcription software technology such as Google’s speech recognition software and Kaldi’s toolkit. The behavior and performance of these two software pieces have been analyzed for different sample frequencies in the range from 300Hz to 44,1kHz.

With widespread use of advanced technology for the recording, storing and sharing of social interactions, protecting privacy of people has been a growing concern. This paper zooms in on the collection of spoken audio with regard for the privacy of recorded individuals. Recently efforts have been made to collect audio at a low sampling rate to obfuscate spoken words in the recorded audio, such that conversations are kept private. This research investigates whether it is possible to upsample this low-resolution audio, using an existing super-resolution model, in order to reveal parts of the previously obfuscated conversations. The performance of the model is measured in terms of the word error rate of automatically generated transcriptions of the upsampled audio. It turns out that it is possible to significantly increase the intelligibility of low resolution privacy-sensitive audio by upsampling. Though the use of the super-resolution model seems to be limited when it comes to revealing significant parts of conversations.

Detecting social interactions through wireless wearable Bluetooth devices is increasing in popularity. Devices use the signal strength to other detected devices to estimate the proximity between people and group them together based on the Dominant set algorithm. Dominant sets are a maximal clique of nodes with an edge-weight based on the affinity between the nodes. Nevertheless, the signal is heavily influenced by external factors, which increase in an crowded environment. This paper introduces three different noise reduction filters that try to detect the kind of noise and therefore improve the detection of surrounded devices. Further, this paper looks at the overall impact of proximity on the resulting RSSI values. Knowing this relationship helps to normalize the values and therefore eliminates the need to apply noise reduction. Using a dataset of 48 sensors recorded in a conference setting with a specific designed sensor the low frequency pass filter gets an accuracy score of 81.8 % with a cut-off frequency of 0.07 Hz. It performs best when considering a conversation window of 20 seconds. Here, only 2/3 of the detected groups has to coincide with the actual formed group at a specific timestamp. Furthermore, the orientation of the participants to each other has heavy influence on the resulting RSSI values and therefore a normalization based on only proximity cannot be done.

In our daily life people encounter many social interactions, for example in the supermarket, at work and in schools. Currently the most reliable way to find social interactions in groups, is to manually annotate the data. Manual annotation takes a lot of time and human resources and as the information stream goes faster and faster, the manual labour cannot keep up. Therefore an automated program to replace this is desirable. One method to automate this, is by looking at the proximity of people, which has been done by Dikker [1]. The results look promising, however they are sensitive to errors. The F1-score (equation 3) is only 0.625 and the precision (equation 5) is 0.696. This means that there are still a lot of false positives. With orientation data one can determine if the people in question are facing each other before they are assigned to be in social interaction. With a view frustum (cone shape) it can be determined who are facing at each other by checking if the cones overlap. This cone can take different sizes, making the view area bigger or smaller. Because the absolute rotation were only used, it was used to find out who cannot face each other. Using the proximity results as a baseline, the impossible orientations were removed. With the goal to reduce the amount of false positives. From the results it can be concluded that a small cone will perform worse compared to the baseline. The bigger the angle of the cone becomes, the closer it gets to the results from Dikker [1]. However, in none of the results it turns out that using absolute orientations improves the current work significantly.

The Midge is a wearable badge created by the Socially Perceptive Computing Lab, Pattern Recognition and Bioinformatics group of the Delft University of Technology, with as goal to analyse human behaviour. The badge has a digital motion processor (DMP) that can determine its orientation. This DMP makes use of an inertial measurement unit (IMU), that houses an accelerometer, a gyroscope and a magnetometer, to calculate its movement in a 3D-space. For both of the accelerometer and gyroscope the Full Scale Range (FSR) can be changed, in addition to the frequency. In this paper, the effects of both elements are analysed to determine if they influence the accuracy of the data gathered. The results show that the changing the FSR does not influence the accuracy of neither the two sensors nor the performance of the DMP. On the other hand, it was found that changing the frequency does influence the performance of the Midge. Even though the frequency did not affect the measurements of the accelerometer and gyroscope directly, the performance of the DMP was affected. The DMP performed best with a frequency of 150 Hz. Using a higher frequency also captured local extremes and turning points from the sensors and the interpreted orientation more precisely.