Guard against silence-based language misdetection

[paberr]

By default the first few seconds of audio are used to detect the language.
However, if you start recording ahead of the meeting, these might be mostly silent and detection fails.

Maybe it should instead use the first seconds above a threshold.