Creating new versions of

device recordings


In the DAPS dataset, we have provided twelve device recordings using two devices and seven acoustic environments. However, it could be useful to create additional device recordings with other device types in other acoustic environments. Additionally, one could create new device recordings using the clean raw data (as opposed to the clean data) in order to preserve non-speech vocal sounds such as breaths and lip smacks in the device recordings. We outline a simple procedure below to create new device recordings using the tools that we provide (available with the dataset) to aid in this process.

1) Concatenate Clean Data - Use the provided Matlab script (concat_data.m) to concatenate all of the clean or clean raw recordings that are to be recorded onto a device (we used all excerpts of all speakers of the clean recordings to create the provided device recordings).

2) Record on Device - Play the concatenated recording through a high quality loudspeaker in a given acoustic environment and record it into a device. In order to better approximate a human mouth, it could be beneficial to use a coaxial loudspeaker (such as the Presonus Sceptre S6 that we used) rather than a two-way or three-way loudspeaker, as it better approximates a point source. The recording application that we used on the tablet and smartphone is Voice Recorder HD. Any application that does not perform any processing and that supports the desired sampling rate should work equally well.

3) Segment and Align Recorded Data - Use the provided Matlab script (organize_rec_data.m) to automatically segment the device recording into individual excerpts from individual speakers. The script also aligns the segmented recordings to the corresponding clean and produced recordings, and save them as wave files.