The use of photoplethysmogram signal (PPG) for heart monitoring is commonly
found nowadays in smartphones and wrist wearables. Besides heart rate or sleep
monitoring common usage, it has been proved that information from PPG can be extracted
for other uses, like person verification, for example. In this work, we evaluate
whether if speech/non-speech events can be inferred from fluctuations they might
cause in the pulse signal. In order to do so, an exploration on end-to-end convolutional
neural ...
The use of photoplethysmogram signal (PPG) for heart monitoring is commonly
found nowadays in smartphones and wrist wearables. Besides heart rate or sleep
monitoring common usage, it has been proved that information from PPG can be extracted
for other uses, like person verification, for example. In this work, we evaluate
whether if speech/non-speech events can be inferred from fluctuations they might
cause in the pulse signal. In order to do so, an exploration on end-to-end convolutional
neural network architectures is done for performing both feature extraction
and classification of the mentioned events. The results are motivating, detecting
speech in PPG signal with a 68.2% AUC using the best performing architecture.
On the other hand, a first experiment on speaker’s voice pitch detection is done, in
order to check if a prosody marker such as pitch variation could be present in PPGs,
but such clue is not clearly found in the results obtained. Nevertheless, the correlation
between speech and PPG signal is proven and the way is paved for further
experiments on this topic.
+