Speech detection
Also known as: voice activity detection, talk detection
Finding where somebody is talking on an audio track, without transcribing the words, which is enough to protect speech from being cut apart.
Updated
What it means
Speech detection answers a smaller question than transcription. It reports the time ranges where a voice is present and nothing about what was said, which turns out to be most of what an editor needs.
Those ranges are the boundaries an automatic cutter has to respect. A cut between phrases is invisible, and a cut through a word is the most noticeable editing error there is, so knowing where words are matters more than knowing which they were.
It is also far cheaper than running a full Transcript, which means it can be applied to every clip on a card rather than to the handful somebody suspects contain talking.
How it works in practice
The output drives two things. Edit points are nudged into the gaps between detected ranges by Auto-cut, and the music level is dropped across them so anything spoken stays intelligible under the track.
A short margin either side of each range is worth adding before Smart trimming uses them. Detectors tend to clip the first consonant and the last, and a tenth of a second of padding keeps whole words intact rather than nearly intact.
Running it on the original camera track rather than an assembled mix is essential, because once music is laid underneath the measurement reports speech almost everywhere.
What to watch out for
Wind defeats it in the worst way, by producing false positives. A gust has energy in the same range as a voice, so a whole ride can read as continuous talking with nothing audible in it.
Background voices are detected as readily as foreground ones. A conversation behind the camera protects itself from being cut while contributing nothing anybody wants to hear.
Shouting over noise is also missed frequently, because a voice at the same level as the Engine sound around it has no distinguishing envelope for a detector to find.
How RawClip handles it
Every clip is checked for speech before cut points are placed, so an edit lands between phrases rather than through them, and the music steps back wherever words were found. The check runs on the original camera audio rather than the finished mix.