Speech Recognition on TIMIT (Latency and Power Metrics)
0.14PER DegradationBlock-Circulant RNN FFT8
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Block-Circulant RNN FFT8Block size=8, Matrix Size (#Params of top layer)=0.41M, Model Compression Ratio=7.9 : 1, Platform=KU0602018.03 | 0.14 | 13.7 | 231,514 | — | — | |
| Block-Circulant RNN FFT8Block size=8, Matrix Size (#Params of top layer)=0.41M, Model Compression Ratio=7.9 : 1, Platform=7V32018.03 | 0.14 | 12.9 | 240,389 | 24 | 10,016 | |
| ESEMatrix Size (#Params of top layer)=0.73M, Model Compression Ratio=4.5 : 1, Platform=KU0602018.03 | 0.3 | 57 | 17,544 | 41 | 428 | |
| Block-Circulant RNN FFT16Block size=16, Matrix Size (#Params of top layer)=0.20M, Model Compression Ratio=15.9 : 1, Platform=KU0602018.03 | 0.31 | 7.4 | 429,327 | — | — | |
| Block-Circulant RNN FFT16Block size=16, Matrix Size (#Params of top layer)=0.20M, Model Compression Ratio=15.9 : 1, Platform=7V32018.03 | 0.31 | 8.3 | 382,510 | 25 | 15,300 |