vocabulary should contain blank to make the decoder API easy to use.

[chenjiasheng]

models/deep_speech_2/model_utils/decoder.py

Line 87 in 8b5c739

raise ValueError("The shape of prob_seq does not match with the "

If vocabulary dosen't contain blank, we have:

vocabulary=['a', 'b', ..., 'z', ' ', '.']
len(vocabulary) = 28
len(prob_list) = 29

Then at line 135,

new_char = vocabulary[c]

here c ranges from 0 to len(prob_list), but vocabulary[28] causes index out of range.

To use this decoder, one have to make the vocabulary a dict indexed from 1 instead of a list indexed from 0, in spite of that the doc suggests vocabulary to be a list.

    :param vocabulary: Vocabulary list.
    :type vocabulary: list

This is really confusing.

@xinghai-sun
@kuke

[kuke]

@chenjiasheng Thanks for your attention!

This wouldn't happen because we in fact have enforced blank_id equal to len(vocabulary). And blank_id shouldn't be passed as an argument to the decoder, we have fixed that in the C++ decoders and will fix the python version very soon.

[chenjiasheng]

@kuke OK, thanks for the reply.