czech data preparation for ASR - ffmpeg with pipe cause get_utt2dur.sh crash

The data preparation for Czech language uses ffmepg together with pipe in wav.scp which does not support wave length in the header.

The wav.scp command is created here: https://github.com/espnet/espnet/blob/a5742a3b23d8a27c0c0ef02d105e1ab9d6321e08/egs/commonvoice/asr1/local/data_prep.pl#L57

The ffmpeg problematic issue is described at: https://trac.ffmpeg.org/ticket/7892

The error in the output of run.sh - the script level:

# script level for run.sh
utils/data/get_utt2dur.sh: could not get utterance lengths from sphere-file headers, using wav-to-duration
run.pl: 4 / 4 failed, log is in data/train_cs/log/get_durations.*.log

The error demonstration when using and not using the pipe with FFmpeg. See the duration info!

The test_bad.wav was redirected to the file (same behavior as used by pipe).

The test_ok.wav was saved to disc by FFmpeg directly. All other FFmpeg parameters are the same.

oplatek@hydra4:master-replicate-czech:asr1$ ffmpeg -i download/cs_data/cv-corpus-5.1-2020-06-22/cs/clips/common_voice_cs_20500128.mp3 -ar 16000 -acodec pcm_s16le -ac 1 -f wav - > test_bad.wav 2> /dev/null                           
oplatek@hydra4:master-replicate-czech:asr1$ ffmpeg -i download/cs_data/cv-corpus-5.1-2020-06-22/cs/clips/common_voice_cs_20500128.mp3 -ar 16000 -acodec pcm_s16le -ac 1 -f wav test_ok.wav 2> /dev/null
oplatek@hydra4:master-replicate-czech:asr1$ soxi test_bad.wav test_ok.wav                                          

Input File     : 'test_bad.wav'
Channels       : 1
Sample Rate    : 16000
Precision      : 16-bit
Duration       : 37:16:57.73 = 2147483647 samples ~ 1.00663e+07 CDDA sectors                                       
File Size      : 96.8k
Bit Rate       : 5.77
Sample Encoding: 16-bit Signed Integer PCM


Input File     : 'test_ok.wav'
Channels       : 1                                                                                                 
Sample Rate    : 16000
Precision      : 16-bit                                                                                            
Duration       : 00:00:03.02 = 48384 samples ~ 226.8 CDDA sectors                                                  
File Size      : 96.8k
Bit Rate       : 256k                                                                                              
Sample Encoding: 16-bit Signed Integer PCM

Total Duration of 2 files: 37:17:00.75

Workaround: save the files to wave first.

Any ideas how to make it work without saving the wave files first?

[oplatek]

FYI: I have managed to run successfully first two stages with the workaround from https://github.com/kaldi-asr/kaldi/pull/4467/files.

Feel free to close this issue. I hit problems with LM training with stage 3 and I moved to the Espnet2 recipe in egs2/commonvoice/asr1 as you suggested - thank you! I hit another issue which I reported here: https://github.com/espnet/espnet/issues/3042

[kamo-naoyuki]

FYI: I used this simple workaround https://github.com/kaldi-asr/kaldi/pull/4467/files and set --read-entire-file to true when I run get_utt2dur.sh on the data before speed perturbation.

Good, this is better for us, thank you!

Btw: I think that FFmpeg could potentially solve it by reading mp3 twice (I assume some CLI option would be needed e.g. --compute-length) or use the information from a header - unfortunately MP3 header does not contain duration information.

Potentially yes, maybe, ffmpeg doesn’t prepare such options, I’m not sure. Another idea is creating a reading tool to do it. This is not difficult.