do your test it in cpu?

[hdmjdp]

whether it can run in realtime?

[YuzukiTsuru]

According to the paper

FFTNet models produce audio more quickly(> 70× faster) than the Fast WaveNet formulation, there by enabling real-time synthesis on a single CPU. Moreover, we show that when used as a vocoder, FFTNet produces higher quality synthetic voices, as measured by a “mean opinion score” test

As you can see, FFTNet is much faster than WaveNet and can be used on the CPU, but there is no mention that it can be synthesized on the CPU in real time.