Training on long audio

[chlorane]

I have tried Whisper to carry ASR task. It performs well.
However, in our case, we want to train and test our model on long audios (about 70sec). Our data are given directly using 40d fbank features and we tried to use some fc layers to convert them to fit the 80d input. Unfortunately, now whisper only seems to allow a 30s audio input, making much information be cut off before being conveyed to the network.
For this case, is there any solution?