Dynamic Ringtone Adjustment using Audio Sampling on Android Smartphones Casey O’Leary – Washington State University Yong Li – University of Tennessee Undergraduate Research Assistant, University of Tennessee Graduate Student Introduction Because the Android operating system is open source, this platform was chosen for the research. The device used for testing was the LG Nexus 4 running a modified version of Android 4.4 KitKat. Cell phones have become an integral part of everyday life in the modern day world. This research looks at the issue of environmental ambient noise that reduces the ability to detect a ringtone during an incoming call to their device. A solution to this problem is sample and analyze the environment, and play pre-determined ringtones that have contrasting frequencies with the ambient noise. The intensity of ambient sounds must also be considered, as the environment may be somewhat silent, or very intense. For this reason, ambient sound intensity is also considered when choosing an appropriate ringtone and volume level for the cell phone. The algorithm implemented and discussed in this research was developed by Christopher Daffron and Alex Hoppe. Sampling Frequency Thresholds In order for the phone to ring, CallNotifier uses the public ring() function from the Ringer class. Within this function, we have added our own function, which runs right before the call is displayed to the user. This function use multiple other functions, which start an audio recording thread, perform the necessary signal processing, and then extract the average frequency and intensity from that data. Cat 1 -> 3: X < 2000 Hertz Cat 2 -> 4: 2000 Hertz < X < 3300 Hertz Cat 3 -> 5: 3300 Hertz < X < 4500 Hertz Cat 4 -> 2: 4500 Hertz < X < 6000 Hertz Cat 5 -> 1: 6000 Hertz < X On Demand Sampling Need environmental characteristics at the moment right before incoming call is displayed Microphone on device used to record raw audio data stored in WAV file at a frequency of 44,100 Hertz A 500 millisecond audio sample provides approximately 22,000 samples between 0 and 20 Kilohertz Fast Fourier Transform Changes raw audio data into a frequency domain Allows for frequency and intensity data to be extracted from data points Frequency and Intensity Calculations Weighted Average frequency of sample; computed by the summation of the intensity at each frequency multiplied by the frequency and then divide that by the summation of all frequency intensity values Average intensity of sample; computed by taking the average value of all of the absolute value of the samples Intensity Thresholds X < .012 (0) .012 < X < .030 (1) .030 < X < .048 (2) .048 < X < .067 (3) .067 < X < .085 (4) .085 < X < .103 (5) .103 < X < .12 (6) X > .14 (7) Source Code Integration The implementation of the sampling and analysis methods requires modifications to the Android operation system, specifically the phone package. Here, we have successfully intercepted the incoming call, taken an audio sample, performed the for mentioned calculations, and set the ringtone and volume. Once the incoming call is detected, a system message is broadcasted. This specific message is received and handled in the CallNotifier class, which contains an instance of the Ringer object. Testing Future Work and Challenges Algorithm Optimization Including recording, sampling, and processing the algorithm has an average runtime of 3709.16 milliseconds Eliminate unnecessary complexity and looping Implement a multi-thread strategy to process data Noise Reduction using Signal Processing Identifying any unwanted “noisy” frequencies that create varying results in processing and remove them before processing Signal Processing via Cloud Computing Use a cloud, such as Amazon EC2, to perform necessary data processing and send results back to device Consider wireless signal strength to determine whether to use cloud or local device This graph demonstrates the need for filtering and algorithm improvement. There is far too much variation in the processing of one single environment, excluding the quiet patio. Most notably, the Market Square Afternoon data shows the inconsistency of the algorithm to consistently categorize an environment’s weighted average frequency of a sample. This work was supported primarily by the Engineering Research Center Program of the National Science Foundation and the Department of Energy under NSF Award Number EEC-1041877 and the CURENT Industry Partnership Program.