[cbruns]

The most efficient way is a filter bank technique. Key words to Google would be "reconstruction", "dechannelizer", "synthesis".

A power of two number (N) of equally spaced frequency channels can be efficiently combined using a short fir filter on each channel followed by an fft where one sample from each channel is input to the fft per frame. Then you get N samples of your output. There is a bit more nuance but the author's bandwidth and number of channels are trivial to handle with this method.

Since each channel is already assumed to be spaced in frequency, you are essentially already in the frequency domain and only so only one fft stage is required.