Data-structure change in split_scp.pl to prevent it becoming very slow when speakers have many utterances.

egs/wsj/s5/utils/split_scp.pl

@@ -100,10 +100,10 @@ if ($utt2spk_file ne "") {  # We have the --utt2spk option...
         if(!defined $spk_count{$s}) {
             push @spkrs, $s;
             $spk_count{$s} = 0;
-            $spk_data{$s} = "";
+            $spk_data{$s} = [];  # ref to new empty array.
         }
         $spk_count{$s}++;
-        $spk_data{$s} = $spk_data{$s} . $_;
+        push @{$spk_data{$s}}, $_;
     }
     # Now split as equally as possible ..
     # First allocate spks to files by allocating an approximately
@@ -182,7 +182,7 @@ if ($utt2spk_file ne "") {  # We have the --utt2spk option...
             $error = 1;
         } else {
             foreach $spk ( @{$scparray[$scpidx]} ) {
-                print F $spk_data{$spk};
+                print F @{$spk_data{$spk}};
                 $count += $spk_count{$spk};
             }
             if($count != $scpcount[$scpidx]) { die "Count mismatch [code error]"; }