Search

We present a new machine learning based bed occupancy detection system that uses only the accelerometer signal captured by a bed-attached consumer smartphone. Automatic bed occupancy detection is necessary for automatic long-term cough monitoring since the time that the monitored patient occupies the bed is required to accurately calculate a cough rate. Accelerometer measurements are more cost-effective and less intrusive than alternatives such as video monitoring or pressure sensors. A 249-hour dataset of manually-labelled acceleration signals gathered from seven patients undergoing treatment for tuberculosis (TB) was compiled for experimentation. These signals are characterised by brief activity bursts interspersed with long periods of little or no activity, even when the bed is occupied. To process them effectively, we propose an architecture consisting of three interconnected components. An occupancy-change detector locates instances at which bed occupancy is likely to have changed, an occupancy-interval detector classifies periods between detected occupancy changes and an occupancy-state detector corrects falsely-identified occupancy changes. Using long short-term memory (LSTM) networks, this architecture achieved an AUC of 0.94. To demonstrate the application of this bed occupancy detection system to a complete cough monitoring system, the daily cough rates along with the corresponding laboratory indicators of a patient undergoing TB treatment were estimated over a period of 14 days. This provides a preliminary indication that automatic cough monitoring based on bed-mounted accelerometer measurements may present a non-invasive, non-intrusive and cost-effective means of monitoring the long-term recovery of patients suffering from respiratory diseases such as TB and COVID-19.

The wood used to make musical instruments needs to have particular properties. Depending on its function, such as a soundboard for string instruments or the body of a wind instrument, different properties are desirable to obtain the best musical quality. Several different classification schemes exist that correlate physical and mechanical properties of wood to define desirable ranges for tonewoods, and to allow suitable wood species to be chosen. The physical and mechanical properties of various wood species indigenous to southern Africa were characterised and then assessed in terms of their suitability for violin construction using these classification schemes. The results of this analysis show that the most suitable of the wood species assessed are yellowwood and sapele. These were subsequently used by a professional luthier to build an ‘African’ violin. The sound quality of this instrument was determined subjectively through performances to an audience and more objectively via spectral analysis of audio recordings. This analysis shows clear differences in the relative magnitude of the harmonics between the violin made from indigenous wood and an instrument made with conventional wood species. Despite the differences, yellowwood and sapele were found to be suitable tonewoods, resulting in an instrument with a unique sound. Significance: • Good quality violins are always made from spruce and maple wood, which have to be imported to South Africa – often at high cost. However, the growth conditions of most southern African wood species should make them suitable tonewoods. We showed that several species are suitable to be used as tonewoods and that the sound produced with such a violin is – although somewhat different – of high quality.

Abstract Dynamic time warping (DTW) can be used to compute the similarity between two sequences of generally differing length. We propose a modification to DTW that performs individual and independent pairwise alignment of feature trajectories. The modified technique, termed feature trajectory dynamic time warping (FTDTW), is applied as a similarity measure in the agglomerative hierarchical clustering of speech segments. Experiments using MFCC and PLP parametrisations extracted from TIMIT and from the Spoken Arabic Digit Dataset (SADD) show consistent and statistically significant improvements in the quality of the resulting clusters in terms of F-measure and normalised mutual information (NMI).

In this study, we present improvements in N-best rescoring of code-switched speech achieved by n-gram augmentation as well as optimised pretraining of long short-term memory (LSTM) language models with larger corpora of out-of-domain monolingual text. Our investigation specifically considers the impact of the way in which multiple monolingual datasets are interleaved prior to being presented as input to a language model. In addition, we consider the application of large pretrained transformer-based architectures, and present the first investigation employing these models in English-Bantu code-switched speech recognition. Our experimental evaluation is performed on an under-resourced corpus of code-switched speech comprising four bilingual code-switched sub-corpora, each containing a Bantu language (isiZulu, isiXhosa, Sesotho, or Setswana) and English. We find in our experiments that, by combining n-gram augmentation with the optimised pretraining strategy, speech recognition errors are reduced for each individual bilingual pair by 3.51% absolute on average over the four corpora. Importantly, we find that even speech recognition at language boundaries improves by 1.14% even though the additional data is monolingual. Utilising the augmented n-grams for lattice generation, we then contrast these improvements with those achieved after fine-tuning pretrained transformer-based models such as distilled GPT-2 and M-BERT. We find that, even though these language models have not been trained on any of our target languages, they can improve speech recognition performance even in zero-shot settings. After fine-tuning on in-domain data, these large architectures offer further improvements, achieving a 4.45% absolute decrease in overall speech recognition errors and a 3.52% improvement over language boundaries. Finally, a combination of the optimised LSTM and fine-tuned BERT models achieves a further gain of 0.47% absolute on average for three of the four language pairs compared to M-BERT. We conclude that the careful optimisation of the pretraining strategy used for neural network language models can offer worthwhile improvements in speech recognition accuracy even at language switches, and that much larger state-of-the-art architectures such as GPT-2 and M-BERT promise even further gains.

Abstract Background The ability to study animal behaviour is important in many fields of science, including biology, behavioural ecology and conservation. Behavioural information is usually obtained by attaching an electronic tag to the animal and later retrieving it to download the measured data. We present an animal-borne behaviour classification system, which captures and automatically classifies three-dimensional accelerometer data in real time. All computations occur on specially designed biotelemetry tags while attached to the animal. This allows the probable behaviour to be transmitted continuously, thereby providing an enhanced level of detail and immediacy. Results The performance of the animal-borne automatic behaviour classification system is presented for sheep and rhinoceros. For sheep, a classification accuracy of 82.40% is achieved among five behavioural classes (standing, walking, grazing, running and lying down). For rhinoceros, an accuracy of 96.10% is achieved among three behavioural classes (standing, walking and lying down). The estimated behaviour was established approximately every 5.3 s for sheep and 6.5 s for rhinoceros. Conclusions We demonstrate that accurate on-animal real-time behaviour classification is possible by successful design, implementation and deployed on sheep and rhinoceros. Since the bandwidth required to transmit the behaviour class is lower than that which would be required to transmit the accelerometer measurements themselves, this system is better suited to low-power and error-prone data communication channels that may be expected in the animals habitat.

Agglomerative hierarchical clustering becomes infeasible when applied to large datasets due to its O(N2) storage requirements. We present a multi-stage agglomerative hierarchical clustering (MAHC) approach aimed at large datasets of speech segments. The algorithm is based on an iterative divide-and-conquer strategy. The data is first split into independent subsets, each of which is clustered separately. Thus reduces the storage required for sequential implementations, and allows concurrent computation on parallel computing hardware. The resultant clusters are merged and subsequently re-divided into subsets, which are passed to the following iteration. We show that MAHC can match and even surpass the performance of the exact implementation when applied to datasets of speech segments.