This talk will survey several recent projects dealing with recorded speech. The first explores 
an optical process for recovering sound recorded onto postcards using a forgotten technology 
from more than a century ago. This involves scanning the postcard at multiple orientations using 
a flatbed scanner, and then reconstructing the fine scale surface texture of the card (where the 
audio is encoded) using photometric stereo, a technique from computer vision. We will then 
discuss more modern applications as well, including a text-based interface for editing recorded 
audio narration that is capable of synthesizing new words matching the voice of the narrator. 
Finally, given that real-world audio recordings are often degraded by factors such as noise, 
reverberation, and equalization distortion, we will also introduce a deep learning method to 
transform recorded speech to sound as though it had been recorded in a studio.