Pocket-sphinx text alignment problem for words having alternative pronunciation

[nayan kalita]

Hi All,

I am using pocketsphinx, ps_alignment to align a text and to get phoneme level timings . For some word, i have multiple alternative pronounciation in the dictionary.

AND AE N D
AND(2) AH N D
THEMSELVES DH EH M S EH L V Z
THEMSELVES(2) DH AH M S EH L V Z

I am using following code for text alignment ,

.....
       al = ps_alignment_init(d2p);
        ps_alignment_add_word(al, dict_wordid(dict, "<s>"),0);
        for(i = 0; i < nwords; i++)
        {   
            ps_alignment_add_word(al, dict_wordid(dict, words[i]),0);

        }
        ps_alignment_add_word(al, dict_wordid(dict, "</s>"),0);
        ps_alignment_populate(al);      
        search = state_align_search_init("state_align",config, acmod, al);
        ....

     acmod_start_utt(acmod);
    ps_search_start(search);
    while (!feof(rawfh)) {
        nread = fread(buf, sizeof(*buf), 2048, rawfh);
        bptr = buf;
        while ((nfr = acmod_process_raw(acmod, &bptr, &nread, FALSE)) > 0) {
            while (acmod->n_feat_frame > 0) {
                ps_search_step(search, acmod->output_frame);
                acmod_advance(acmod);
            }
        }
    }
    ps_search_finish(search);
        ...

It is always aligned to the first pronunciation of the word for all words. How I can solve this problem?

[Nickolay V. Shmyrev]

First decode to figure out exact pronunciation then align to get time boundaries.

[nayan kalita]

Hi Nickolay,

Thanks for your suggestion.