CTM from transcript file and audio file

[Mikel Esparza]

Hi!

I would like to get the CTM file given an audio file and a transcript file from that audio. I already have prepared and tested, acoustics models (one GMM model and a NNet model) and language models (n-grams and lstm models).

I've read that GMM models works better for aligning transcript files and audio files. For my task I started from segment_long_utterances.sh file and I changed some code to transform the CTM_edits file into a final CTM where the insertions and the silences are cleaned.

It already works well with some audios, but I've seen that with some others I get problems in the aligning task. Some transcripts are not well cleaned and the decode_segmentation.sh file fails.

I want to know if there is some magical kaldi recipe which does the task of aligning taking care about all the possible problems.  Or maybe there is a better script to start with rather than segment_long_utterances.sh.

If not, which value of "beam" and "lattice_beam" is enough to force an output  from the decode_segmentation file? I've already tried --beam=20.0 --lattice-beam=6.0 and it still fails at some points.

Thanks in advance!!

[Daniel Povey]

It's possible that all you need is steps/get_train_ctm.sh.

--
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/4a1e7dea-b825-43fa-9220-648af4eaec01%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Mikel Esparza]

Thanks for your quick response! I have tried the script steps/get_train_ctm.sh  and it works fine!

Now I have problems with long audio files. The script  steps/nnet3/align.sh fails if the audio is too large. what should I do to be able to handle long audio files?

Thanks in advance.

[Daniel Povey]

How does it fail?

--

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/CABLbLk3Utw7p36q9mWPykEjpfOH_QOEsf36YCmcNTUY33RbF9g%40mail.gmail.com.

[Mikel Esparza]

Hi Dan, thank you for your quick response.

I get this error

LOG (nnet3-align-compiled[5.5]:CheckAndFixConfigs():nnet-am-decodable-simple.cc:294) Increasing --frames-per-chunk from 50 to 51 to make it a multiple of --frame-subsampling-factor=3
WARNING (nnet3-align-compiled[5.5]:AlignUtteranceWrapper():decoder-wrappers.cc:501) Retrying utterance AudioCortesAragon1-00002500-00005500-1 with beam 40
WARNING (nnet3-align-compiled[5.5]:AlignUtteranceWrapper():decoder-wrappers.cc:510) Did not successfully decode file AudioCortesAragon1-00002500-00005500-1, len = 825
WARNING (nnet3-align-compiled[5.5]:AlignUtteranceWrapper():decoder-wrappers.cc:501) Retrying utterance AudioCortesAragon1-00010000-00013000-2 with beam 40
WARNING (nnet3-align-compiled[5.5]:AlignUtteranceWrapper():decoder-wrappers.cc:510) Did not successfully decode file AudioCortesAragon1-00010000-00013000-2, len = 365
WARNING (nnet3-align-compiled[5.5]:AlignUtteranceWrapper():decoder-wrappers.cc:501) Retrying utterance AudioCortesAragon1-00047500-00050500-1 with beam 40
WARNING (nnet3-align-compiled[5.5]:AlignUtteranceWrapper():decoder-wrappers.cc:510) Did not successfully decode file AudioCortesAragon1-00047500-00050500-1, len = 1000
LOG (compile-train-graphs[5.5]:main():compile-train-graphs.cc:147) compile-train-graphs: succeeded for 19 graphs, failed for 0
LOG (apply-cmvn[5.5]:main():apply-cmvn.cc:81) Copied 19 utterances.
LOG (nnet3-align-compiled[5.5]:main():nnet3-align-compiled.cc:198) Overall log-likelihood per frame is 4.60517 over 12988 frames.
LOG (nnet3-align-compiled[5.5]:main():nnet3-align-compiled.cc:201) Retried 3 out of 19 utterances.
LOG (nnet3-align-compiled[5.5]:main():nnet3-align-compiled.cc:203) Done 16, errors on 3
LOG (nnet3-align-compiled[5.5]:~CachingOptimizingCompiler():nnet-optimize.cc:710) 0.103 seconds taken in nnet3 compilation total (breakdown: 0.0323 compilation, 0.0541 optimization, 0 shortcut expansion, 0.0114 checking, 0.000498 computing indexes, 0.00514 misc.) + 0 I/O.

Since I'm not going to train any model with this data, I would like to align the full transcript text with the audio file.

Thanks in advance.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CAEWAuySoc_tgZsrYc0e%2Bh7G7ECNuoP4g6mu86b6-Ccheyq5S%2Bw%40mail.gmail.com.

[Daniel Povey]

You could try increasing the beam and retry-beam.  It could indicate a problem with the data or transcripts though.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CABLbLk3Tmah7KCYYBDkdC8E0CmbH9DKHjR7xinrL%3DwYragoRkw%40mail.gmail.com.

[Mikel Esparza]

Analyzing the data I've realized that there are 2 minutes of silence in the audio. To try to solve this, I've thought to use a SAD model to subtract the silence from the audio. For that, I've used a pretrained SAD model. Now I have a segments file which contain only the speech parts.

The problem is that I have the transcript of the full audio file. Is there a way to tell Kaldi to align the full transcript file but only with those parts where the SAD model has said there is speech?

The only solution I see is to create a new audio file, cropping those silence parts and align that with the full transcript, but maybe there is a way to specify the same to Kaldi with metadata.

Thanks in advance!

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CAEWAuyQmNM%3DYDmeyuDAWQ8-K6c5sMqU9JgmRLchXtNzg7K1NTA%40mail.gmail.com.

[Daniel Povey]

I don't know of an easy way to do that.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CABLbLk2Q%3DJzs8QUrWKfBpTTFBe6bvc%3DehpwF%3DFf7A94ND59BsA%40mail.gmail.com.

[Vimal Manohar]

You can use segment_long_utterances.sh giving the whole recording level transcript as text for the segments. You can do this manually or using <text-in> <utt2text> arguments.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CABLbLk2Q%3DJzs8QUrWKfBpTTFBe6bvc%3DehpwF%3DFf7A94ND59BsA%40mail.gmail.com.

[Mikel Esparza]

It worked perfectly fine!

Thank you so much

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CAGLZPU-x6ZeeC1jR%3DBk8gUn7p9N9-bt6yobkgt7ezNyEPRdsGQ%40mail.gmail.com.