That ran in 4.194 on my dataset, which can be reduced by simplifying the regex even more.

open my $io, "<", "test.csv";
open my $oh, ">", "out.csv";

while (<$io>) {
    s{ ("[^""]+") }{ (my $one = $1) =~ tr{,}{-}; $one =~ tr{""}{}d; $o
+ne; }xge;
    print $oh $_;
    }
[download]

runs in 3.229. All regex-based scripts will fail if

• the first field is quoted;
• the second field has a embedded double-quote (or an escaped character with the default " as escape)
• any record anywhere in the dataset has an embedded newline, and the data after the newline has a double-quote

As long as you are absolutely certain that the CSV data is uniformly and consistently laid out as in these two lines, you are safe.

I would personally never take that risk, unless that two seconds are a problem. 5 seconds for 1.4 mln records is pretty fast, knowing it is always safe.