| 摘要 |
<p>A speech recognition system (in eg. in a mobile phone) stores an input signal from a microphone (100 or 102) in a first buffer 110 and generates a noise-reduced signal 134 in a second block 146 before selecting 140 which of the signals to send to a speech recognition engine 132. Aspects of the invention include the detection 120 and validation 130 of a trigger phrase or pass phrase (eg. hello phone) which, when spoken by an authorised speaker (identified by speaker verification), wakes up the speech recognition engine, with a time delay being applied to either signal stored in buffers 110 and 146 in order to synchronise them. This multi-phase process (fig. 4) allows computationally intensive speech recognition to be automatically powered down between uses.</p> |
