One word produced by :word_timestamps.
Times are absolute seconds within the input audio. probability is the
minimum per-token acoustic probability across the tokens that make up
this word (matching faster-whisper's per-word confidence reduction);
filter at e.g. probability < 0.3 to flag low-confidence words.