besacier / mboshi-french-parallel-corpus Goto Github PK

Hi,

I've found that there some forced alignments files in forced_alignments_supervised_spkr/all that do not match any audio files in full_corpus_newsplit/all. I think it would be better to remove them as it's easier to prepare the features/alignments for any ASR toolkit.

Lucas

Hi,

It seems that this utterance is complete silence:
abiayi_2015-09-09-13-14-50_samsung-SM-T530_mdw_elicit_Dico10_197

Best,
Lucas

Thank you for sharing the low-resource parallel corpus with alignment and its French translation.

I'm trying to use some phonological knowledge on the Mboshi, but I encountered some confusability of telling which IPA phone corresponding to the annotated phone in full_corpus_newsplit.

Mboshi has 68 different phones，however in full_corpus_newsplit corpus there are only 31 phones for annotation, I'm trying hard to create the mapping from the 68 phones to the 31 phones, but it seems to be hard.

I am wondering is it possible to opensource the mapping file from 68 phones to the 31 phones, or where else can I get the information of the phone mapping?

I have checked these two papers but there is not enough information to get the mapping
"A Very Low Resource Language Speech Corpus for Computational Language Documentation Experiments"
"Developing an Embosi (Bantu C25) Speech Variant Dictionary to Model Vowel Elision and Morpheme Deletion"

Hi Laurent,

There are some audio wav files that are corrupted. Most likely the length stored in some wav files does not match the actual number of samples. This may affect some softwares, for instance HCopy (from HKT) will fail to extract features for those corrupted files. I've managed to fix the problem by using the following:

    $ cat fix_wav.sh
    #!/bin/bash
    wav="$1"
    outdir="$2"
    bname=$(basename $wav)
    sox --ignore-length "$wav" "$outdir/$bname"
    $ mkdir mboshi-french-parallel-corpus/full_corpus_newsplit/all.fixed
    $ find mboshi-french-parallel-corpus/full_corpus_newsplit/all -name '*wav' -exec ./fix_wav.sh {} mboshi-french-parallel-corpus/full_corpus_newsplit/all.fixed \;

I was then able to use HTK to extract the features for all the files.

Best,
Lucas