Difference between revisions of "QSLocal-history"

From cslt Wiki
Jump to: navigation, search
 
(22 intermediate revisions by 3 users not shown)
Line 1: Line 1:
 +
 
{| class="wikitable"
 
{| class="wikitable"
!FreeNeb Release List!!!!                                              !!!!          !!          !!          !!        !!                                                            !!            !!
+
!FreeNeb Release List!!                                              !!          !!           
 
|-
 
|-
!  Date    !!!!      Release Name                                    !!!! Category || Project ||  Version !!Relesaer !!            Location                                        ||  Document  || History ||
+
!  Date    ||     Version       || note    ||QR Code
 
|-
 
|-
|2018-01-05 ||||Chinese ASR Local |||| APP    || internal ||  v1.0  ||  ZMY        || http://srv1.freeneb.com/static/shareFiles/ASR/asr-local.apk ||  https://gitlab.com/freeneb/demo-project        ||  ||
+
|   || README || 时实更新的README文档,使用前请先认真阅读 || [https://gitlab.com/freeneb/app-vpr-local/blob/master/doc/README]
 +
|-
 +
|12.14||V4.2.1||更新log:在x-vector模型中,加入了nnet-vad,并把nnet模型压缩至int16。引擎更新到2.1.1 模型更新到2.1.4|| [[Image:VPR_V4.2.1.png|100px|VPR_V4.2.1]]
 +
|-
 +
|12.05||V4.2.0||引擎更新到2.0.1 模型更新到2.1.1 支持x-vector || [[Image:VPR_V4.2.0.png|100px|VPR_V4.2.0]]
 +
|-
 +
|11.20||V4.1.2||Network:  ReLU tdnn, splice (-2,-1,0,1,2 -2,2 0 -1,1 0 -2,2 0 -4,4 0 0) * 1000. Training: trained on VoxCeleb (1+2) for 3 epoches, then full-info adaptation with the same training set for 100 iters. Config:   full-info_fix_inter100_time1_momen0.9 num-gpu-initial=2 num-gpu-final=4 initial_effective_lrate=0.00015, final_effective_lrate=0.00015 LDA:      the same training set.  Add nnet-vad nnet_vad_threshold=-3|| [[Image:VPR_V4.1.2.png|100px|VPR_V4.1.2]]
 +
|-
 +
|10.08||V4.1.1||Network:  ReLU tdnn, splice (-2,-1,0,1,2 -2,2 0 -1,1 0 -2,2 0 -4,4 0 0) * 1000. Training: trained on VoxCeleb (1+2) for 3 epoches, then full-info adaptation with the same training set for 100 iters. Config:   full-info_fix_inter100_time1_momen0.9 num-gpu-initial=2 num-gpu-final=4 initial_effective_lrate=0.00015, final_effective_lrate=0.00015 LDA:      the same training set.  Add nnet-vad nnet_vad_threshold=-3|| [[Image:VPR_V4.1.1.png|100px|VPR_V4.1.1]]
 +
|-
 +
|9.30||V4.1.0||Network:  ReLU tdnn, splice (-2,-1,0,1,2 -2,2 0 -1,1 0 -2,2 0 -4,4 0 0) * 1000. Training: trained on VoxCeleb (1+2) for 3 epoches, then full-info adaptation with the same training set for 100 iters. Config:  full-info_fix_inter100_time1_momen0.9 num-gpu-initial=2 num-gpu-final=4 initial_effective_lrate=0.00015, final_effective_lrate=0.00015 LDA:      the same training set. || [[Image:VPR_V4.1.0.png|100px|VPR_V4.1.0]]
 +
|-
 +
|9.25||V3.11.0||Network:  ReLU tdnn, splice (-2,-1,0,1,2 -2,2 0 -1,1 0 -2,2 0 -4,4 0 0) * 1000. Training: trained on VoxCeleb (1+2) for 3 epoches, then full-info adaptation with the same training set for 100 iters. Config:   full-info_fix_inter100_time1_momen0.9 num-gpu-initial=2 num-gpu-final=4 initial_effective_lrate=0.00015, final_effective_lrate=0.00015 LDA:     the same training set. || [[Image:VPR_V3.11.0.png|100px|VPR_V3.11.0]]
 +
|-
 +
|8.5||V3.10.0||QSLocal 3.10 增加阿里训练数据、新vad模型 || [[Image:VPR_V3.10.0.png|100px|VPR_V3.10.0]]
 +
|-
 +
|7.6||V3.8.5||使用VPR0.6版本引擎  || [[Image:VPR_V3.8.5.png|100px|VPR_V3.8.5]]
 +
|-
 +
|6.29||V3.8.4||使用3.8模型+DNN Vad,VPR0.5版本引擎  || [[Image:VPR_V3.8.4.png|100px|VPR_V3.8.4]]
 +
|-
 +
|6.21||V3.9.3||在3.9版模型基础上,加入了dnn-based vad  || [[Image:VPR_V3.9.3.png|100px|VPR_V3.9.3]]
 +
|-
 +
|6.8||V3.8.3||在V3.8.2增强版 基础上,fix data buffer bug  || [[Image:VPR_V3.8.3.png|100px|VPR_V3.8.3]]
 +
|-
 +
|6.5||V3.8.2增强版||在V3.8.2 基础上,去掉尾Null数据(梦原把这一feature移到了Android-dev),设置default abs threshold=17.0  || [[Image:VPR_V3.8.2.png|100px|VPR_V3.8.2]]
 +
|-
 +
|6.5||V3.8.2||在V3.8.2 基础上,去掉语音前后的context padding || [[Image:VPR_V3.8.2.png|100px|VPR_V3.8.2]]
 +
|-
 +
|6.4||V3.8.1||在V3.8 model基础上,加入SVM VAD || [[Image:VPR_V3.8.1.png|100px|VPR_V3.8.1]]
 +
|-
 +
|6.4||V3.9.1||在v3.9基础上,去掉语音前后的context padding || [[Image:VPR_V3.9.1.png|100px|VPR_V3.9.1]]
 +
|-
 +
|6.1||V3.9||修改模型 训练数据:基于ali-phase2数据 在7500人的full-info模型上 进一步full-info得到的模型|| [[Image:VPR_V3.9.png|100px|VPR_V3.9]]
 +
|-
 +
|5.28||V3.8||修改模型 训练数据:speech-ocean_datatang_7500_mix_clean_reverb_volume_noise;模型:ReLU 预先训练 3 个 epoches,然后用 full-info 自适应 200 个 iterations.修改最后传输数据为空值的bug|| [[Image:VPR_V3.8.png|100px|VPR_V3.8]]
 +
|-
 +
|5.17||V3.7||修复AudioRecord bug,将环形buffer改成线性。 || [[Image:VPR_V3.7.png|100px|VPR_V3.7]]
 +
|-
 +
|5.7||V3.6||石颖大模型:reverbe+volume+noise+clean 总量为基本数据的3又三分之一倍 || [[Image:VPR_V3.6.png|100px|VPR_V3.6]]
 +
|-
 +
|4.27||V3.5||小模型:仅添加noise 总量为基本数据的一半,人数不变 || [[Image:VPR_V3.5.png|100px|VPR_V3.5]]
 +
|-
 +
|4.24 ||V3.4||修改bug || [[Image:VPR_V3.4.png|100px|VPR_V3.4]]
 +
|-
 +
|4.16 ||V3.3||修改滚动样式 || [[Image:VPR_V3.4.png|100px|VPR_V3.3]]
 +
|-
 +
|4.3||V3.0||依时间认证版本 || [[Image:VPR_V3.4.png|100px|VPR_V3.0]]
 +
|-
 +
|3.31||v2.0||带反馈的认证版本 || [[Image:VPR_V2.0.png|100px|VPR_V2.0]]
 +
|-
 +
|3.25||V1.0||盲认证版本|| [[Image:VPR_V1.0.png|100px|VPR_V1.0]]
 
|}
 
|}

Latest revision as of 02:00, 14 December 2018

FreeNeb Release List
Date Version note QR Code
README 时实更新的README文档,使用前请先认真阅读 [1]
12.14 V4.2.1 更新log:在x-vector模型中,加入了nnet-vad,并把nnet模型压缩至int16。引擎更新到2.1.1 模型更新到2.1.4 VPR_V4.2.1
12.05 V4.2.0 引擎更新到2.0.1 模型更新到2.1.1 支持x-vector VPR_V4.2.0
11.20 V4.1.2 Network: ReLU tdnn, splice (-2,-1,0,1,2 -2,2 0 -1,1 0 -2,2 0 -4,4 0 0) * 1000. Training: trained on VoxCeleb (1+2) for 3 epoches, then full-info adaptation with the same training set for 100 iters. Config: full-info_fix_inter100_time1_momen0.9 num-gpu-initial=2 num-gpu-final=4 initial_effective_lrate=0.00015, final_effective_lrate=0.00015 LDA: the same training set. Add nnet-vad nnet_vad_threshold=-3 VPR_V4.1.2
10.08 V4.1.1 Network: ReLU tdnn, splice (-2,-1,0,1,2 -2,2 0 -1,1 0 -2,2 0 -4,4 0 0) * 1000. Training: trained on VoxCeleb (1+2) for 3 epoches, then full-info adaptation with the same training set for 100 iters. Config: full-info_fix_inter100_time1_momen0.9 num-gpu-initial=2 num-gpu-final=4 initial_effective_lrate=0.00015, final_effective_lrate=0.00015 LDA: the same training set. Add nnet-vad nnet_vad_threshold=-3 VPR_V4.1.1
9.30 V4.1.0 Network: ReLU tdnn, splice (-2,-1,0,1,2 -2,2 0 -1,1 0 -2,2 0 -4,4 0 0) * 1000. Training: trained on VoxCeleb (1+2) for 3 epoches, then full-info adaptation with the same training set for 100 iters. Config: full-info_fix_inter100_time1_momen0.9 num-gpu-initial=2 num-gpu-final=4 initial_effective_lrate=0.00015, final_effective_lrate=0.00015 LDA: the same training set. VPR_V4.1.0
9.25 V3.11.0 Network: ReLU tdnn, splice (-2,-1,0,1,2 -2,2 0 -1,1 0 -2,2 0 -4,4 0 0) * 1000. Training: trained on VoxCeleb (1+2) for 3 epoches, then full-info adaptation with the same training set for 100 iters. Config: full-info_fix_inter100_time1_momen0.9 num-gpu-initial=2 num-gpu-final=4 initial_effective_lrate=0.00015, final_effective_lrate=0.00015 LDA: the same training set. VPR_V3.11.0
8.5 V3.10.0 QSLocal 3.10 增加阿里训练数据、新vad模型 VPR_V3.10.0
7.6 V3.8.5 使用VPR0.6版本引擎 VPR_V3.8.5
6.29 V3.8.4 使用3.8模型+DNN Vad,VPR0.5版本引擎 VPR_V3.8.4
6.21 V3.9.3 在3.9版模型基础上,加入了dnn-based vad VPR_V3.9.3
6.8 V3.8.3 在V3.8.2增强版 基础上,fix data buffer bug VPR_V3.8.3
6.5 V3.8.2增强版 在V3.8.2 基础上,去掉尾Null数据(梦原把这一feature移到了Android-dev),设置default abs threshold=17.0 VPR_V3.8.2
6.5 V3.8.2 在V3.8.2 基础上,去掉语音前后的context padding VPR_V3.8.2
6.4 V3.8.1 在V3.8 model基础上,加入SVM VAD VPR_V3.8.1
6.4 V3.9.1 在v3.9基础上,去掉语音前后的context padding VPR_V3.9.1
6.1 V3.9 修改模型 训练数据:基于ali-phase2数据 在7500人的full-info模型上 进一步full-info得到的模型 VPR_V3.9
5.28 V3.8 修改模型 训练数据:speech-ocean_datatang_7500_mix_clean_reverb_volume_noise;模型:ReLU 预先训练 3 个 epoches,然后用 full-info 自适应 200 个 iterations.修改最后传输数据为空值的bug VPR_V3.8
5.17 V3.7 修复AudioRecord bug,将环形buffer改成线性。 VPR_V3.7
5.7 V3.6 石颖大模型:reverbe+volume+noise+clean 总量为基本数据的3又三分之一倍 VPR_V3.6
4.27 V3.5 小模型:仅添加noise 总量为基本数据的一半,人数不变 VPR_V3.5
4.24 V3.4 修改bug VPR_V3.4
4.16 V3.3 修改滚动样式 VPR_V3.3
4.3 V3.0 依时间认证版本 VPR_V3.0
3.31 v2.0 带反馈的认证版本 VPR_V2.0
3.25 V1.0 盲认证版本 VPR_V1.0