Voice wake-up method and device

• Main Authors: WANG FEI, WANG HUANLIANG, WU TIANXIN, XIAO JIALIN
• Format: Patent
• Language: Chinese; English
• Published: 14.04.2023
• Subjects: ACOUSTICS; MUSICAL INSTRUMENTS; PHYSICS; SPEECH ANALYSIS OR SYNTHESIS; SPEECH OR AUDIO CODING OR DECODING; SPEECH OR VOICE PROCESSING; SPEECH RECOGNITION

The invention provides a voice wake-up method and device. The method comprises the following steps: receiving an audio sequence; performing audio event detection on the audio sequence; inputting an audio clip with voice into a first-stage wake-up model; when the first-stage wake-up model detects a wake-up word, inputting the audio clip into a second-stage wake-up model; when the second-level wake-up model detects a wake-up word, inputting the audio clip into the voiceprint model; when the voiceprint model judges that the speaker is a registrant, the system is awakened; wherein the first-stage wake-up model is a streaming model and is modeled by taking a word as a unit; the second-stage wake-up model is a non-streaming model, and attention weights are distributed to a convolution output channel of the second-stage wake-up model. According to the method, the power consumption is reduced by constructing the two-stage wake-up model, the first-stage wake-up model adopts word modeling, so that the second-stage wake