Transly is a sequence to sequence Bi-directional LSTM Encoder-Decoder model that's trained on the CMU pronouncing dictionary, IIT Bombay English-Hindi Parallel Corpus and IIT Kharagpur transliteration corpus.
The pronunciation module in Transly can predict pronunciation of any given word (with an American accent of course!)
Take any word of any language - just transliterate the word in English (all capitals) and you are good to go. Be it a new or old, seen or unseen, sensible or insensible word - Transly can catch'em all!
Another module in Transly is the transliteration module. It currently supports Hindi to English and English to Hindi transliterations.
Pre-trained models can be found inside the respective trained_models folders. New models can also be trained on custom data.
Use the package manager pip to install transly
pip install translyUsing the pre-trained pronunciation model
import transly.pronunciation as tp
# let's try a hindi word
# the prediction accent would be American
QUERY = 'MAKAAN'
a = tp.load_model(model_path='cmu')
a.infer(QUERY, separator=" ")Training a new model on custom data
from transly.seq2seq.config import SConfig
from transly.seq2seq.version0 import Seq2Seq
config = SConfig(training_data_path=training_data_path, input_mode='character_level', output_mode='word_level')
s2s = Seq2Seq(config)
s2s.fit()
s2s.save_model(path_to_model=model_path, model_file_name=model_file_name)Using the pre-trained model
import transly.transliteration as tl
QUERY = 'नहीं'
a = tl.load_model(model_path='hi2en')
a.infer(QUERY)Using the pre-trained model
import transly.transliteration as tl
QUERY = 'NAHI'
a = tl.load_model(model_path='en2hi')
a.infer(QUERY)Training a new model on custom data
from transly.seq2seq.config import SConfig
from transly.seq2seq.version0 import Seq2Seq
config = SConfig(training_data_path=training_data_path)
s2s = Seq2Seq(config)
s2s.fit()
s2s.save_model(path_to_model=model_path, model_file_name=model_file_name)Training data file should be a csv with two columns, the input and the output
| Input | Output |
|---|---|
| AA | AA1 |
| AABERG | AA1 B ER0 G |
| AACHEN | AA1 K AH0 N |
| AACHENER | AA1 K AH0 N ER0 |
The Python code in this module is distributed with Apache License 2.0