Let SSMs be ConvNets: State-space Modeling with Optimal Tensor Contractions

Benchmark Model Rank Results
speech-enhancement-on-demandCentaurus (0.51M)–PESQ (wb): 3.25
speech-recognition-on-librispeech-test-cleanCentaurus (30 M)–Word Error Rate (WER): 4.4
speech-recognition-on-speech-commandsCentaurus–Accuracy (%): 98.53