Keyword Spotting on Google Speech Commands (test)
98AccuracyBC-ResNet8
Evaluation results
| Method | Links | ||
|---|---|---|---|
| BC-ResNet80-shot=false, #Params=312k, #FLOPs=89.1M2026.03 | 98 | — | |
| res15 w/ SSN#Params=243K, S=4, A=Sub, Training Protocol=Same epoch as [5]2021.03 | 97.5 | — | |
| MatchBoxNet0-shot=false, #Params=93k, #FLOPs=11.3M2026.03 | 97.5 | — | |
| MHAtt-RNN0-shot=false, #Params=743k, #FLOPs=22.7M2026.03 | 97.2 | — | |
| PCOV-KWS-M0-shot=true, #Params=376k, #FLOPs=1.8M2026.03 | 96.9 | — | |
| EdgeSpeechNet-A#Params=107K2021.03 | 96.8 | — | |
| res15 w/ SSN#Params=243K, S=4, A=Sub2021.03 | 96.8 | — | |
| PhonMatchNet0-shot=true, #Params=655k2026.03 | 96.8 | — | |
| TC-ResNet14-1.5#Params=305K2021.03 | 96.6 | — | |
| TENet120-shot=false, #Params=100k, #FLOPs=2.9M2026.03 | 96.6 | — | |
| TC-ResNet0-shot=false, #Params=305k, #FLOPs=6.7M2026.03 | 96.6 | — | |
| PCOV-KWS-S0-shot=true, #Params=211k, #FLOPs=1.13M2026.03 | 96.6 | — | |
| S4D-LinModel size=Large, paradigm=Unconstrained Bidirectional (Bi.)2022.06 | 96.25 | — | |
| S4D-InvModel size=Large, paradigm=Unconstrained Bidirectional (Bi.)2022.06 | 96.18 | — | |
| S4-LegSModel size=Large, paradigm=Unconstrained Bidirectional (Bi.)2022.06 | 96.08 | — | |
| S4D-LegSModel size=Large, paradigm=Unconstrained Bidirectional (Bi.)2022.06 | 95.83 | — | |
| res15 w/ BN#Params=239K2021.03 | 95.8 | — | |
| ResNet-150-shot=false, #Params=238k, #FLOPs=894M2026.03 | 95.8 | — | |
| Att-RNN0-shot=false, #Params=202k, #FLOPs=22.3M2026.03 | 95.6 | — | |
| DS-CNNResource Budget=L, Target Memory=500KB, Target Operations=80MOps, Actual Memory=497.6KB, Actual Operations=56.9M, Weight Precision=8-bit2017.11 | 95.4 | — | |
| res8 w/ SSN#Params=113K, S=4, A=Sub2021.03 | 95.4 | — | |
| S4-FouTModel size=Large, paradigm=Unconstrained Bidirectional (Bi.)2022.06 | 95.27 | — | |
| CRNNResource Budget=L, Target Memory=500KB, Target Operations=80MOps, Actual Memory=499.5KB, Actual Operations=19.3M, Weight Precision=8-bit2017.11 | 95 | — | |
| DS-CNNResource Budget=M, Target Memory=200KB, Target Operations=20MOps, Actual Memory=189.2KB, Actual Operations=19.8M, Weight Precision=8-bit2017.11 | 94.9 | — | |
| LSTMResource Budget=L, Target Memory=500KB, Target Operations=80MOps, Actual Memory=498.8KB, Actual Operations=48.4M, Weight Precision=8-bit2017.11 | 94.8 | — | |
| GRUResource Budget=L, Target Memory=500KB, Target Operations=80MOps, Actual Memory=499.7KB, Actual Operations=48.4M, Weight Precision=8-bit2017.11 | 94.7 | — | |
| DS-CNNResource Budget=S, Target Memory=80KB, Target Operations=6MOps, Actual Memory=38.6KB, Actual Operations=5.4M, Weight Precision=8-bit2017.11 | 94.4 | — | |
| CRNNResource Budget=M, Target Memory=200KB, Target Operations=20MOps, Actual Memory=199.8KB, Actual Operations=7.6M, Weight Precision=8-bit2017.11 | 94.4 | — | |
| GRUResource Budget=M, Target Memory=200KB, Target Operations=20MOps, Actual Memory=200.0KB, Actual Operations=19.2M, Weight Precision=8-bit2017.11 | 94.2 | — | |
| res8 w/ BN#Params=111K2021.03 | 94.1 | — | |
| CRNNResource Budget=S, Target Memory=80KB, Target Operations=6MOps, Actual Memory=79.7KB, Actual Operations=3.0M, Weight Precision=8-bit2017.11 | 94 | — | |
| JointBackbone=TC-ResNet-8, Parameters=64.48K, Memory=1624.4M2025.12 | 94 | — | |
| LSTMResource Budget=M, Target Memory=200KB, Target Operations=20MOps, Actual Memory=198.6KB, Actual Operations=19.2M, Weight Precision=8-bit2017.11 | 93.9 | — | |
| S4-LegSModel size=Large, paradigm=Autoregressive (AR)2022.06 | 93.6 | — | |
| S4D-LegSModel size=Large, paradigm=Autoregressive (AR)2022.06 | 93.57 | — | |
| GRUResource Budget=S, Target Memory=80KB, Target Operations=6MOps, Actual Memory=78.8KB, Actual Operations=3.8M, Weight Precision=8-bit2017.11 | 93.5 | — | |
| Basic LSTMResource Budget=L, Target Memory=500KB, Target Operations=80MOps, Actual Memory=494.5KB, Actual Operations=47.9M, Weight Precision=8-bit2017.11 | 93.4 | — | |
| S4D-InvModel size=Large, paradigm=Autoregressive (AR)2022.06 | 93.4 | — | |
| S4D-LinModel size=Large, paradigm=Autoregressive (AR)2022.06 | 93.37 | — | |
| Basic LSTMResource Budget=M, Target Memory=200KB, Target Operations=20MOps, Actual Memory=196.5KB, Actual Operations=18.9M, Weight Precision=8-bit2017.11 | 93 | — | |
| LSTMResource Budget=S, Target Memory=80KB, Target Operations=6MOps, Actual Memory=79.5KB, Actual Operations=3.9M, Weight Precision=8-bit2017.11 | 92.9 | — | |
| CNNResource Budget=L, Target Memory=500KB, Target Operations=80MOps, Actual Memory=497.8KB, Actual Operations=25.3M, Weight Precision=8-bit2017.11 | 92.7 | — | |
| CNNResource Budget=M, Target Memory=200KB, Target Operations=20MOps, Actual Memory=199.4KB, Actual Operations=17.3M, Weight Precision=8-bit2017.11 | 92.2 | — | |
| Basic LSTMResource Budget=S, Target Memory=80KB, Target Operations=6MOps, Actual Memory=63.3KB, Actual Operations=5.9M, Weight Precision=8-bit2017.11 | 92 | — | |
| S4-FouTModel size=Large, paradigm=Autoregressive (AR)2022.06 | 91.78 | — | |
| CNNResource Budget=S, Target Memory=80KB, Target Operations=6MOps, Actual Memory=79.0KB, Actual Operations=5.0M, Weight Precision=8-bit2017.11 | 91.6 | — | |
| CSDNumber of training domains=10002020.03 | 91.3 | — | |
| Rainbow KeywordsBackbone=TC-ResNet-8, Memory size L=3000, Parameters=129.96K, Memory=97.2M2025.12 | 91.3 | -0.012 | |
| CGNumber of training domains=10002020.03 | 91.2 | — | |
| ERMNumber of training domains=10002020.03 | 90.8 | — | |
| Rainbow KeywordsBackbone=TC-ResNet-8, Memory size L=1500, Parameters=129.96K, Memory=48.6M2025.12 | 88.7 | -0.012 | |
| CSDNumber of training domains=2002020.03 | 87.5 | — | |
| RWalkBackbone=TC-ResNet-8, Parameters=129.96K, Memory=178.6M2025.12 | 87.1 | -0.045 | |
| CGNumber of training domains=2002020.03 | 86.9 | — | |
| ERMNumber of training domains=2002020.03 | 86.8 | — | |
| DNNResource Budget=L, Target Memory=500KB, Target Operations=80MOps, Actual Memory=496.6KB, Actual Operations=990.2K, Weight Precision=8-bit2017.11 | 86.7 | — | |
| DNNResource Budget=M, Target Memory=200KB, Target Operations=20MOps, Actual Memory=199.4KB, Actual Operations=397.0K, Weight Precision=8-bit2017.11 | 86.4 | — | |
| DNNResource Budget=S, Target Memory=80KB, Target Operations=6MOps, Actual Memory=80.0KB, Actual Operations=158.8K, Weight Precision=8-bit2017.11 | 84.6 | — | |
| iCaRLBackbone=TC-ResNet-8, Parameters=75.29K, Memory=178.6M2025.12 | 84.6 | -0.057 | |
| PCL-KWSBackbone=TC-ResNet-8, Parameters=406.9K, Memory=162.4M2025.12 | 83.6 | -0.041 | |
| EWCBackbone=TC-ResNet-8, Parameters=129.96K, Memory=162.4M2025.12 | 83.5 | -0.064 | |
| Rainbow KeywordsBackbone=TC-ResNet-8, Memory size L=500, Parameters=129.96K, Memory=16.2M2025.12 | 82.8 | -0.036 | |
| CSDNumber of training domains=1002020.03 | 81.4 | — | |
| CGNumber of training domains=1002020.03 | 80.4 | — | |
| ERMNumber of training domains=1002020.03 | 80 | — | |
| BiCBackbone=TC-ResNet-8, Parameters=64.48K, Memory=178.6M2025.12 | 79.3 | -0.085 | |
| CSDNumber of training domains=502020.03 | 73.7 | — | |
| CGNumber of training domains=502020.03 | 73.3 | — | |
| ERMNumber of training domains=502020.03 | 72.6 | — | |
| NRBackbone=TC-ResNet-8, Parameters=64.48K, Memory=178.6M2025.12 | 56 | -0.163 | |
| Fine-tuneBackbone=TC-ResNet-8, Parameters=64.48K, Memory=162.4M2025.12 | 26.2 | -0.372 |