Detectors work from two kinds of evidence.
Timing and sound.Classic AMD measures the audio at the start of the call. Asterisk's AMD application, for example, returns MACHINE when the silence before the greeting, the greeting itself or its word count runs past set limits, and HUMAN when the greeting is followed by enough silence. MightyCall's AMD page lists voice tone and pitch, the duration of pauses and background noise as its signals, and Call Center Studio's says it analyzes voice and silence durations.
Words.Speech-based detection transcribes the greeting and looks for voicemail wording, such as “leave a message” or “is not available”.
VoiceOI works from the words, and recorded-voice scripts can add a greeting-length check.