Fix mozilla#693; corrections for several Fisher test samples

bin/import_fisher.py

@@ -42,8 +42,14 @@ def _download_and_preprocess_data(data_dir):
     # their audio source. The result is that we end up with more labels than time
     # slices, which breaks CTC.
     all_2004.loc[all_2004["wav_filename"].str.endswith("fe_03_00265-33.53-33.81.wav"), "transcript"] = "correct"
+    all_2004.loc[all_2004["wav_filename"].str.endswith("fe_03_00991-527.39-528.3.wav"), "transcript"] = "that's one of those"
     all_2005.loc[all_2005["wav_filename"].str.endswith("fe_03_10282-344.42-344.84.wav"), "transcript"] = "they don't want"
+    all_2005.loc[all_2005["wav_filename"].str.endswith("fe_03_10677-101.04-106.41.wav"), "transcript"] = "uh my mine yeah the german shepherd pitbull mix he snores almost as loud as i do"
 
+    # The following file is just a short sound and not at all transcribed like provided.
+    # So we just exclude it.
+    all_2004 = all_2004[~all_2004["wav_filename"].str.endswith("fe_03_00027-393.8-394.05.wav")]
+
     # The following file is far too long and would ruin our training batch size.
     # So we just exclude it.
     all_2005 = all_2005[~all_2005["wav_filename"].str.endswith("fe_03_11487-31.09-234.06.wav")]