A speech recognition apparatus and a method are provided in which even when a speech recognition grammar is employed independently or alternatively, speech recognition response is improved. Voice data is received. Then, an output suspended state for a speech recognition result is maintained until the duration of a silence interval that follows the utterance part reaches a criterion time. Information indicating whether a last word of a word sequence is a final word is stored. On the basis of the language model, a recognition candidate word sequence is extracted. When the last word of the extracted word sequence is determined as being a final word, the speech recognition result is output in a time shorter than the criterion time, while when determined as not being a final word, the speech recognition result is output at the time that the criterion time has elapsed.