The present disclosure discloses a voice interaction method, system, terminal device and medium, with operations performing voice recognition on collected voice signals to acquire an input sentence; semantically matching the input sentence with cached sample sentences, determining whether there is a sample sentence having same or similar semantics as the input sentence among the cached sample sentences; if yes, acquiring cached response content having the same or similar semantics as the input sentence; if not, sending at least one of the input sentence or the collected voice signals to a server; receiving response content or the collected voice signals, as response content of the input sentence, acquired through semantic understanding according to a knowledge base, responding to the input sentence according to the response content of the input sentence, and updating at least one of the cached sample sentences or the response content of the cached sample sentences.