This paper proposes exactly what the title says: a new method for modified discrete cosine transforms (MDCTs). However, there is a problem with the abstract (and indeed throughout the text). The abstract claims that the method applies for “N=2 × qm; m ≥ 2 where q is an odd integer” (and that further factors of 2 can be added), but the practical examples given are for N being 12, 36 or 640, with only 36 having m ≥ 2. In fact, I cannot see where m ≥ 2 is used in the paper.
N being 12 or 36 comes from MPEG 1/2, and 640 comes from G729.1 (ITU). Table 6 unfortunately shows that, for these three cases, there is an already-published algorithm (a different one in each case) whose operation count is no worse than, and for 12 and 640 strictly better than, the proposed new algorithm. The authors claim that their algorithm “has the merits of parallelism and numerical stability,” but I fail to see any justification for these claims.