36 percent is correct.
The time taken without TLB will be $2 \times T_m$, as once we'll have to look for the correct address in the memory and one more access for actually accessing that value.
The time taken with TLB will be $T_{\text{tlb} + T_m}$ as we'll get the address from the TLB and then we can access the memory to find the relevant data.
Going by that, the respective times are 1000 ns and 640 ns.
The speedup will be given by: $\frac{\text{Old Value - New Value}}{\text{Old Value}} \times 100$ which will be $\frac{1000-640}{1000} \times 100$ which is 36 %.