CMU Sphinx 5prealpha alignment issue

Question

I am using sphinx 5prealpha to do alignment but I am getting bad results. I tried different AM and dictionnaries the results is always the same. When I use the same AM and dictionaries with an older version (sphinx4) I get very good result. for this audio file and text : files

The result is good for sphinx4 and not for 5prealpha. I am sure if you try with any french AM and dictionary you will see the difference. Is there any way to fix that issue. Any help will be appreciated.

---- UPDATE ----

I tried with the two AM and dictionaries here : Ester and SPhinx models

I am using this source code: Sphinx source code

When I use ester AM and dictionary, I expect to get (result from the old sphinx version 4): expected.txt

But I get (with 5 prealpha : what_i_get.txt

I am displaying result with :

                        List<WordResult> wr = aligner.align(audioUrl, text);

                    for (WordResult result : wr) {

                        System.out.println(
                                result.getWord().toString() + " " + Long.toString(result.getTimeFrame().getStart())
                                        + "  " + Long.toString(result.getTimeFrame().getEnd()));
                    }

You need to provide result on what exactly do you get from alignment and what do you expect. Also which exactly model do you use. Here it aligns fine except minor issues. — Nikolay Shmyrev
@NikolayShmyrev I updated my question. Thank you in advance for your help. — user2576975

Nikolay Shmyrev Nikolay Shmyrev · Accepted Answer · 2016-09-11T20:19:24

Old algorithm worked best for short utterances like yours, new one is designed for very long files. So new algorithm is not that great for short utterances.

One has to fix slightly the alignment algorithm to make it work for case. I'm planning to look on it this week, but probably it will take some more time.

CMU Sphinx 5prealpha alignment issue

1 Answers