A stretch of a video's sound as 16 kHz mono floats in -1, 1, what speech models take: decoded by the phone's audio decoder, channels averaged, resampled linearly.
fromMs to fromMs + lengthMs; empty if the video has no sound.