Add wav2vec2 ASR Spanish pretrained model from voxpopuli

[mthrok]

Add Spanish ASR from Voxpopuli.

[carolineechen]

not a big deal since this is used internally only, but since we are adding several languages, thoughts on having a dictionary mapping lang -> symbols and using generic function _get_labels(lang) instead of a having a different function for each added language?

[mthrok]

tl;dr sure we can do that.

---

My original intention here was to have the separate label object instance object for different pipeline, (meaning, id function would report different values for each get_labels) because labels would be global object and modifying one would not affect the others.

However, now that get_labels returns tuple, which is immutable, and is constructed at runtime, it is no longer relevant. so we can do that.

[mthrok]

hmm, so I tried adopting the idea, and I think having them as separate function is better for the readability.

The follwing is only for French, but each language will have around 30 lines. So at the end the dictionary will be around 13 * 30 lines.

def _get_voxpopuli_labels(lang):
    labels = {
        'fr': (
            "|",
            "e",
            "s",
            "n",
            "i",
            "t",
            "r",
            "a",
            "o",
            "u",
            "l",
            "d",
            "c",
            "p",
            "m",
            "é",
            "v",
            "q",
            "f",
            "g",
            "b",
            "h",
            "x",
            "à",
            "j",
            "è",
            "y",
            "ê",
            "z",
            "ô",
            "k",
            "ç",
            "œ",
            "û",
            "ù",
            "î",
            "â",
            "w",
            "ï",
            "ë",
            "ü",
            "æ",
        )
    }
    return labels[lang]

[mthrok]

@nateanl

I realized that the last dimension of the ASR output label is 1, which only happened once in the original training dataset. I believe this is a sort of mistake, and we can exclude such dimension from the model like the case of <pad> (#1914).

What do you think?

cat dict.es_char.txt
| 1518049
e 1077702
a 877225
o 694767
s 619500
n 565808
r 519910
i 513128
l 393015
d 382689
c 359006
t 358015
u 323059
p 236137
m 235987
b 88865
q 85869
y 78375
g 77028
v 67765
h 62163
ó 58918
f 53486
í 36803
á 30274
j 25768
z 22749
ñ 19116
é 16605
x 11350
ú 7163
k 1741
w 484
ü 256
1 1        <<--