# US 2019/0378529 A1 · Voice processing method, apparatus, device and storage medium > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 2019/0378529 A1 - Document ID: US20190378529A1 - Status: Published application - Kind: A1 - Application number: 16/236,261 - Filed: 2018-12-28 - Published: 2019-12-12 - Earliest claimed: 2018-06-11 - Assignees: Baidu Online Network Technology (Beijing) Co., Ltd. - Applicants: Baidu Online Network Technology (Beijing) Co., Ltd. - Inventors: CHAO LI; WEIXIN ZHU - Web page: https://patentgrep.com/patent/US20190378529A1 ## CPC - G10L 21/0208: Noise filtering - G06N 3/08: Learning methods - G06N 3/09: Supervised learning - G10L 21/0272: Voice signal separating - G10L 25/30: using neural networks - G10L 25/84: for discriminating voice from noise - G10L 2021/02087: {the noise being separate speech, e.g. cocktail party} - G10L: SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS ## Abstract The present application provides a voice processing method, an apparatus, a device, and a storage medium, including: acquiring a first acoustic feature of each of N voice frames, where N is a positive integer greater than 1; applying a neural network algorithm to N first acoustic features to obtain a first mask; modifying the first mask according to VAD information of the N voice frames to obtain a second mask; and processing the N first acoustic features according to the second mask to obtain a second acoustic feature, resulting in more effective noise suppression and a lower damage to the voice. ## Claims ### Claim 1 A voice processing method, comprising: acquiring a first acoustic feature of each of N voice frames, wherein N is a positive integer greater than 1; applying a neural network algorithm to N first acoustic features to obtain a first mask; modifying the first mask according to voice activity detection (VAD) information of the N voice frames to obtain a second mask; and processing the N first acoustic features according to the second mask to obtain a second acoustic feature. ### Claim 2 (depends on 1) The method according to claim 1, wherein the modifying the first mask according to the VAD information of the N voice frames comprises: calculating a product of the VAD information and the first mask to obtain the second mask. ### Claim 3 (depends on 1) The method according to claim 1, wherein the VAD information comprises a VAD value corresponding to each of the voice frames; and when the N voice frames comprise a silence frame, setting a VAD value corresponding to the silence frame to zero. ### Claim 4 (depends on 1) The method according to claim 1, wherein the VAD information comprises a VAD value corresponding to each of the voice frames; and correspondingly, before the modifying the first mask according to VAD information of the N voice frames, the method further comprises: determining M1 voice frames having a VAD value of 1 and P1 voice frames having a VAD value of 0 from the N voice frames, wherein the M1 voice frames are adjacent to the P1 voice frames, and wherein M1 and P1 are positive integers greater than 1; and smoothing the VAD value corresponding to M2 voice frames of the M1 voice frames and the VAD value corresponding to P2 voice frames of the P1 voice frames, such that the VAD value corresponding to the M2 voice frames and the VAD value corresponding to the P2 voice frames are changed gradually from 0 to 1 or from 1 to 0, wherein the M2 voice frames are adjacent to the P2 voice frames, and wherein 1≤M2≤M1, 1≤P2≤P1. ### Claim 5 (depends on 4) The method according to claim 4, wherein the determining the M1 voice frames having the VAD value of 1 and the P1 voice frames having the VAD value of 0 from the N voice frames comprise: determining a call type corresponding to each of the N voice frames, wherein the type comprises silence and phone; determining a voice frame having the silence type as a voice frame having the VAD value of 0; and determining a voice frame having the phone type as a voice frame having the VAD value of 1. ### Claim 6 (depends on 5) The method according to claim 5, wherein M2 and P2 are determined by a hamming window, a triangular window or a hanning window. ### Claim 7 A voice processing device, comprising: a memory and a processor; wherein the memory is configured to store instructions executed by the processor, such that the processor performs: acquiring a first acoustic feature of each of N voice frames, wherein N is a positive integer greater than 1; applying a neural network algorithm to N first acoustic features to obtain a first mask; modifying the first mask according to voice activity detection (VAD) information of the N voice frames to obtain a second mask; and processing the N first acoustic features according to the second mask to obtain a second acoustic feature. ### Claim 8 (depends on 7) The device according to claim 7, wherein the processor is configured to: calculate a product of the VAD information and the first mask to obtain the second mask. ### Claim 9 (depends on 7) The device according to claim 7, wherein the VAD information comprises a VAD value corresponding to each of the voice frames; and correspondingly, the processor is further configured to: when the N voice frames comprise a silence frame, set a VAD value corresponding to the silence frame to zero. ### Claim 10 (depends on 7) The device according to claim 7, wherein the VAD information comprises a VAD value corresponding to each of the voice frames; and correspondingly, the processor is further configured to: determine M1 voice frames having a VAD value of 1 and P1 voice frames having a VAD value of 0 from the N voice frames, wherein the M1 voice frames are adjacent to the P1 voice frames, and wherein M1 and P1 are positive integers greater than 1; and smooth the VAD value corresponding to M2 voice frames of the M1 voice frames and the VAD value corresponding to P2 voice frames of the P1 voice frames, such that the VAD value corresponding to the M2 voice frames and the VAD value corresponding to the P2 voice frames are changed gradually from 0 to 1 or from 1 to 0, wherein the M2 voice frames are adjacent to the P2 voice frames, and wherein 1≤M2≤M1, 1≤P2≤P1. ### Claim 11 (depends on 10) The device according to claim 10, wherein the processor is configured to: determine a call type corresponding to each of the N voice frames, wherein the type comprises silence and phone; determine a voice frame having the silence type as a voice frame having the VAD value of 0; and determine a voice frame having the phone type as a voice frame having the VAD value of 1. ### Claim 12 (depends on 11) The device according to claim 11, wherein M2 and P2 are determined by a hamming window, a triangular window or a hanning window. ### Claim 13 A non-volatile storage medium, comprising: computer executable instructions for implementing following steps: acquiring a first acoustic feature of each of N voice frames, wherein N is a positive integer greater than 1; applying a neural network algorithm to N first acoustic features to obtain a first mask; modifying the first mask according to voice activity detection (VAD) information of the N voice frames to obtain a second mask; and processing the N first acoustic features according to the second mask to obtain a second acoustic feature. ### Claim 14 (depends on 13) The non-volatile storage medium according to claim 13, wherein the computer executable instructions are configured to implement: calculating a product of the VAD information and the first mask to obtain the second mask. ### Claim 15 (depends on 13) The non-volatile storage medium according to claim 13, wherein the VAD information comprises a VAD value corresponding to each of the voice frames; and correspondingly, the computer executable instructions are further configured to implement: when the N voice frames comprise a silence frame, setting a VAD value corresponding to the silence frame to zero. ### Claim 16 (depends on 13) The non-volatile storage medium according to claim 13, wherein the VAD information comprises a VAD value corresponding to each of the voice frames; and correspondingly, the computer executable instructions are further configured to implement: determining M1 voice frames having a VAD value of 1 and P1 voice frames having a VAD value of 0 from the N voice frames, wherein the M1 voice frames are adjacent to the P1 voice frames, and wherein M1 and P1 are positive integers greater than 1; and smoothing the VAD value corresponding to M2 voice frames of the M1 voice frames and the VAD value corresponding to P2 voice frames of the P1 voice frames, such that the VAD value corresponding to the M2 voice frames and the VAD value corresponding to the P2 voice frames are changed gradually from 0 to 1 or from 1 to 0, wherein the M2 voice frames are adjacent to the P2 voice frames, and wherein 1≤M2≤M1, 1≤P2≤P1. ### Claim 17 (depends on 16) The non-volatile storage medium according to claim 16, wherein the computer executable instructions are configured to implement: determining a call type corresponding to each of the N voice frames, wherein the type comprises silence and phone; determining a voice frame having the silence type as a voice frame having the VAD value of 0; and determining a voice frame having the phone type as a voice frame having the VAD value of 1. ### Claim 18 (depends on 17) The non-volatile storage medium according to claim 17, wherein M2 and P2 are determined by a hamming window, a triangular window or a hanning window.