mohamad-hasan-sohan-ajini / g2p

Grapheme To Phoneme

Attention Based Grapheme To Phoneme

The G2P algorithm is used to generate the most probable pronunciation for a word not contained in the lexicon dictionary. It could be used as a preprocess of text-to-speech system to generate pronunciation for OOV words.

Dependencies

The following libraries are used:
pytorch
tqdm
matplotlib

Install dependencies using pip:

pip3 install -r requirements.txt

Dataset

Currently the following languages are supported:

1. EN: English
2. FA: Farsi
3. RU: Russian

You could easily provide and use your own language specific pronunciatin doctionary for training G2P. More details about data preparation and contribution could be found in resources.
Feel free to provide resources for other languages.

Attention Model

Both encoder-decoder seq2seq model and attention model could handle G2P problem. Here we train attention based model. The encoder model get sequence of graphemes and produces states at each timestep. Encoder states used during attention decoding. The decoder attends to appropriate encoder state (according to its state) and produces phonemes.

Train

To start training the model run:

LANGUAGE=[EN|FA|RU] python train.py

You can also use tensorboard to check the training loss:

tensorboard --logdir log --bind_all

Training parameters could be found at config.py.

Test

To get pronunciation of a word:

# RU example
python test.py --word питон
u0.d.a0.lj.o1.n.k.a0.<eos>
# EN example
python test.py --word PYTHON
P.IH.TH.AH.N.<eos>
# FA example
python test.py --word پایتون
p.A.y.t.o.n.<eos>

You could also visualize the attention weights, using --visualize:

EN	FA	RU

Other implementations

Also check Persian_G2P, for encoder-decoder seq2seq model implementation.

A C++ implementation that relies on trained models of this repository, could be found here.

g2p's Issues

some cases the model can not produce any outputs.

I trained a model based on different English Database. when i run to predict phonemes of my test Database.
In some words (Graphemes), the model stuck and can not produce any output. I think it is because the RNN.
Do you know how can i solve this issue?

Merging with Tihu

Can you merge this into Tihu project?
I am frequently receiving email (mostly from the non-Persian that wants to learn Persian) and they are complaining about G2P module in Tihu. It is old and not working working with Python 3>.

It will be great if you can merge this G2P into Tihu.

License?

Hello,

I'm very interested in your code and the possibilites of using an attention-based g2p model. Could you give me any information on which license your code is available under and in particular whether there are any restrictions on using it?
Thank you very much in advance.

Regards,
M.G.