Decoding Error: Assertion Error

[carlo...@edu.unito.it]

Hi everyone,

i have monophone and triphone system, everything is fine but i have errors in the decoding part of my run.sh script.

the same error is reported in https://groups.google.com/forum/#!msg/kaldi-help/w07niwViMI8/InYIXiqzBAAJ

Here is my case for the triphone (it'sthe same for monophone) :

steps/decode.sh --config conf/decode.config --nj 1 --cmd run.pl exp/tri1/graph data/test exp/tri1/decode
decode.sh: feature type is delta
steps/diagnostic/analyze_lats.sh --cmd run.pl exp/tri1/graph exp/tri1/decode
run.pl: job failed, log is in exp/tri1/decode/log/analyze_alignments.log
steps/decode.sh: Not scoring because local/score.sh does not exist or not executable.

i opened the log file and it reports the following:

Traceback (most recent call last):
  File "steps/diagnostic/analyze_phone_length_stats.py", line 179, in <module>
    assert num_utterances > 0
AssertionError

I am looking at the python script and at its inputs which are contained in the phone_stats.*.gz file:

1 all 6 224
1 all 6 240
1 begin 6 224
1 begin 6 240 

none of them are negative. What could be a possible solution?

Thanks in advance.

[Daniel Povey]

I think that assertion in analyze_lats.sh relates to a bug that has now been fixed.

Regarding:

not scoring because local/score.sh does not exist or not executable.

it means what it says.

------------------------

Indirizzo istituzionale di posta elettronica degli studenti e dei laureati dell'Università degli Studi di Torino

Official University of Turin email address for students and graduates 

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To post to this group, send email to kaldi...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/aaa2d652-6f35-4cd3-a9ee-6cb5d4391d00%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[carlo...@edu.unito.it]

So what would you suggest to solve the problem? my kaldi install is from last week

[carlo...@edu.unito.it]

i re-installed kaldi and tried again, but the error seems to remain. Also, decode has acceso to score.sh because it is specified in the PATH variable

i upload a zip with my example folder for reference

[Daniel Povey]

It needs to be called local/score.sh, that script invokes it with a path.  It would usually be a soft link to ../steps/score_kaldi.sh.

That bug is fixed in the latest version of Kaldi, you would just need to do "git pull" or re-download.  Your zipfile contains the old version.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/1bedad69-84c8-4ed8-a7a8-4549071ce3ee%40googlegroups.com.

[carlo...@edu.unito.it]

i have no problem with score.sh, i am able to run it and get the WER.

My problem is with the  steps/diagnostic/analyze_phone_length_stats.py script, i have an assertion error, as stated in the log file subl exp/tri1/decode/log/analyze_alignments.log :

# gunzip -c exp/tri1/decode/phone_stats.*.gz | steps/diagnostic/analyze_phone_length_stats.py exp/tri1/graph
# Started at Wed Dec 19 15:43:25 CET 2018
#

Traceback (most recent call last):

  File "steps/diagnostic/analyze_phone_length_stats.py", line 171, in <module>
    assert num_utterances > 0
AssertionError
# Accounting: time=0 threads=1
# Ended (code 1) at Wed Dec 19 15:43:25 CET 2018, elapsed time 0 seconds

Anyway i re-upload a version of the example with the scoring working, for reference.

[Daniel Povey]

The bug is in the script that calls it.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/6dd317ab-8d10-43c7-9a84-7a2cf0512c65%40googlegroups.com.

[carlo...@edu.unito.it]

I couldn't figure out what the problem is, if you could be more specific it would be appreciated.

[Daniel Povey]

I told you already:

"That bug is fixed in the latest version of Kaldi, you would just need to do "git pull" or re-download.  "

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/9baaac32-3155-4c24-9bb2-7ad0d057fd30%40googlegroups.com.

[carlo...@edu.unito.it]

Dear Daniel,

i still haven't fixed my problem yet. I reinstalled kaldi ultiple times but  keep having the same error.

I updated everything i could but i still get the num utterances error in my steps/diagnostic/analyze_

phone_length_stats.py

The error is in this part of the program:

# This block will print warnings if silence is seen less than 80% of the time at utterance
# beginning and end.
for boundary_type in 'begin', 'end':
    phone_to_lengths = phone_lengths[boundary_type]
    num_utterances = total_phones[boundary_type]
    assert num_utterances > 0
    opt_sil_lengths = phone_to_lengths[optional_silence_phone]
    frequency_percentage = sum(opt_sil_lengths.values()) * 100.0 / num_utterances
    # The reason for this warning is that the tradition in speech recognition is
    # to supply a little silence at the beginning and end of utterances... up to
    # maybe half a second.  If your database is not like this, you should know;
    # you may want to mess with the segmentation to add more silence.
    if frequency_percentage < 80.0:
        print("analyze_phone_length_stats.py: WARNING: optional-silence {0} is seen only {1}% "
              "of the time at utterance {2}.  This may not be optimal.".format(
                optional_silence_phone_text, frequency_percentage, boundary_type))

Also, i always get a 100% WER, i guess the error could be icaused by this error in the decoding process

[Daniel Povey]

It seems to be failing because nothing was decoded.

Look in your exp/make_mfcc/train/make_mfcc_train.1.log , there are problems there (missing files, too many channels).

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/e0c3459d-81c7-430b-8f55-9c43636a74ee%40googlegroups.com.

[Syun]

And you'd better check the lang/G.fst. 

I printed the lat.1.gz(via lattice-copy command) and noticed that many phones didn't exist in the L.fst(phones like 107 and 94), so I checked the HCLG.fst and G.fst(I used "fstprint" comand to print these fsts), and then I found the G.fst was empty.

Shin

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/e0c3459d-81c7-430b-8f55-9c43636a74ee%40googlegroups.com.

[Muhammad Talha Umair Alim]

I have the same problem, many phone don't exist in L.fst. So, what should I do now in this regard? Thank You

[Muhammad Talha Umair Alim]

Hi Syun,
 after finding that your G.fst was empty then what did you do after that?

On Friday, 25 January 2019 at 06:24:34 UTC+5 Syun wrote: