Text normalization in speech recognition tasks

When solving tasks related to speech recognition (Speech-To-Text) and generation (Text-To-Speech) of speech, it is important that the transcript matches what the speaker said - that is, real spoken language. This means that before written speech becomes our transcript, it needs to be normalized .

In other words, the text needs to be done through several steps:

• Replacing the number in words: 1984 -> one thousand nine hundred and eighty-fourth year ;
• Explanation of abbreviations: 2 . -> two minutes of hatred ;
• Latin transcription: Orwell-> etc.

In this article, I will briefly talk about how normalization developed in the dataset of the Russian language Open_STT , what tools were used and about our approach to the task.

Like a cherry on a cake, we decided to put our normalizer based on seq2seq in the public domain: a link to github . It is as simple as possible to use and is called by one method:

norm = Normalizer()
result = norm.norm_text(' 9  11   whiskas')

>>> '      '

More about the task

, , ? , . , .

, , :

• 2 — (), . 2 — ;
• 2 — , 2 — ;
• 2 — , 2 — ;
• = 2/5 — , . 2/5 — — .

: - ( — ) ( — ?). , , . .

. - . , , ~20% , ~80% .

Open_STT : — . STT , - 2020 , .

. 2020 . "" — .

Sequence to Sequence

- , sequence-to-sequence (seq2seq) . , seq2seq , "" , :

• — ;
• — ;
• , — ;

attention "5 ". , "" "5", "".

seq2seq PyTorch . — . + + + , — + .

— , . ( , ) .

, :

• — Russian Text Normalization;
• -, .;
• , — , , ..

TorchScript

, , Torchscript.

TorchScript — PyTorch, Python C++.

, PyTorch :

1. , TorchScript , ;
2. torch.jit.script ( torch.jit.trace), .

, , , , . , : , .. .

, . , .

• norm.norm_string(" — — 27 38 %.")

' — — .'

• norm.norm_string(" 22 1939 ")

' '

• norm.norm_string(" «The Crying Game»")

' « »'

• norm.norm_string(" XVIII ")

' '

• norm.norm_string(" 2012 6,6 ")