I would not agree on that.

__In terms of sound quality:__
FFT (frequency domain) convolution can be very close to direct (time domain) convolution, and even more precise for very long FIRs. This all goes down to the implementation, but good one do exist.

Now if you compare direct convolution with IIR biquads the latter is much more prone to quantization errors due to the recursion nature, and of course the longer the ringing (lower frequency and/or higher Q) the more the errors. Direct convolution it is the most straightforward and simple (brutal!) way of applying coefficients, with a simple multiplication and addition, and quantization errors are minimal.

__In terms of latency:__
Direct convolution does not imply any latency: one sample goes in, one sample goes out. Now the FIR itself can imply a latency, and this is typically the case with linear-phase correction, but this is not necessarily the case. If the FIR represents a minimum-phase correction (like what IIRs have to stick to) then it can be made zero latency as well.

This is an important and alas often overlooked thing: IIR can only be causal (minimum-phase or excess phase), whereas FIR can be acausal (linear-phase, etc.), *but* can also be causal (minimum-phase or excess phase). In fact FIR can do anything when it comes to magnitude/phase relation, whereas IIR is limited by causality.

Time domain (FFT) convolution processing adds its own latency (FFT buffer), but this can be much reduced using partitioning (like what BruteFIR does), and even zeroed by using direct convolution for the first part of the FIR.

__In terms of processing power:__
Direct convolution is easy for a DSP or an FPGA, but pretty difficult for a computer CPU.

FFT convolution is very easy for a CPU, but care has to be taken on the implementation, both in term of sound quality and processing latency.

__In terms of portability / predictability:__
IIR/biquad implementations are all over the place when it comes to the actual response you get for a given EQ and filter setting.

Not even talking about EQ convention here (proportional Q vs constant Q, shelving cuttoff frequency position, etc. This is a mess), but simple application of a given set of biquad coefficients. Depending on the implementation the response will vary from one device/software to another, and you cannot pass settings around and expect results to be identical, even when you can be sure both devices use the same conventions.

On the other hand a FIR literately represents the transfer function, and if the convolution process is properly implemented it will always give the same result regardless of the device/software that apply it. Direct or FFT convolution all the same, WYSIWYG!

Me personally I like direct convolution the best, and only use FIR even for purely minimum-phase corrections.

I like the brutal nature of this most simple algorithm

my .2€