README.md

metadata

license: mit
language:
  - pt
pretty_name: Male/Female dataset for forced alignment
size_categories:
  - n<1K

MF: Male/Female for Forced Phonetic Alignment

This is a two-speaker dataset manually aligned at the phoneme level (Praat's textgrid files provided). Each portion contains 200 instances from both a male and a female speaker. This repo has been used as ground truth for the following papers:

• Souza and Neto, PROPOR 2016
• Dias et al., BRACIS 2020
• Batista and Neto. BRACIS 2021
• Batista and Neto. PROPOR 2022
• Batista et al., EURASIP 2022

One major problem with both male and female datasets is that they were hand aligned according to an unknown phoneset, which nobody really knows where it actually came from. So it required some considerable amount of processing in order to convert the original phoneset to the FalaBrasil phoneset, since the latter is the one which all acoustic models are trained over.

Dataset original phone set:

$ find male/textgrid -name "*.TextGrid" > filelist
$ while read line ; do awk '/phones/,/syll/' $line | grep 'text =' | awk '{print $3}' ; done < filelist | sort | uniq | sed 's/"//g'
  4    6    6~   a    a~   b    d    e    e~   E    f    g    h    h/   h\   i    i~   j    j~   J
  k    l    L    m    n    o    o~   O    p    s    S    t    u    u~   v    w    w~   z    Z    _

FalaBrasil original phone set (see also our NLP Generator library, which performs the G2P conversion):

$ find male/textgrid_falabrasil -name "*.TextGrid" > filelist
$ while read line ; do awk '/"phones"/,/"syllphones"/' $line | grep 'text =' | awk '{print $3}' ; done < filelist | sort | uniq | sed 's/"//g'
  a    a~   b    d    dZ   e    e~   E    f    g    i    i~   j    j~   J    k    l    L    m    n
  o    o~   O    p    r    R    s    S    t    tS   u    u~   v    w    w~   X    z    Z    _