Phonetic dictionary provides system the data to map vocabulary words to sequence of phonemes. It looks like this:
hello H EH L OW world W ER L D
Dictionary can contain alternative pronunciations, in that case you can designate them with a number in parenthesis
the TH IH the(2) TH AH
There are various phonesets to represent phones like IPA or SAMPA, CMUSphinx does not yet require you to use any well-known phoneset, moreover, it prefers to use letter-only phone names without special symbols. This requirement simplifies some processing algorithms, for example, you can create files with phone names.
Dictionary should contain all the words you are interested in, otherwise recognizer will not be able to recognize them. However, it is not sufficient to have the words in the dictionary, the recognizer looks for the word both in the dictionary and in the language model. Without language model the word will not be recognized even if you added it in the dictionary.
There is no need to remove unused words from the dictionary unless you want to save memory, extra words in the dictionary do not affect accuracy.
There are number of dictionaries which cover languages we support - CMUDict for US English, French, German, Russian, Dutch, Italian, Spanish, Mandarin. Other dictionaries might be found on the web. If dictionary has proper format you can use it.
If dictionary does not cover all the words you are interested in you can extend it with g2p tool.
We recommend to use our latest too g2p-seq2seq . It is based on neural networks implemented in Tensorflow framework and provides a state of the art accuracy of conversion.
An English model 2-layer LSTM with 256 hidden units is available for download on cmusphinx website. Unpack the model after download. It is trained on CMU English dictionary. Read my lips - this model works only for English. For other languages you need to bootstrap dictionary first as described below and then use G2P tool to extend it.
The easiest way to check how the tool works is to run it the interactive mode with model above and type the words
g2p-seq2seq --interactive --model model_folder_path
> hello HH EH L OW
To generate pronunciations for an English word list with a trained model, run
g2p-seq2seq --decode your_wordlist --model model_folder_path
The wordlist is a text file with words, one word per line.
To train G2P you need a dictionary (word and phone sequence per line in standard form). To run the training
g2p-seq2seq --train train_dictionary.dic --model model_folder_path
For more information on the tool see the corresponding page.
If you do not have dictionary for your language there are usually several ways on how you can obtain them.
Usually dictionaries are bootstrapped with hand-written rules. You can find a list of phonemes for your language in Wikipedia page about your language and write a simple Python script to map words to phonemes. The best dictionary could not be covered with rules though, most languages have quite irregular pronunciation which might not be very obvious for newcomer even if it is conventionally thought you speak what is written. This is due to coarticulation effects in human speech. But for basic dictionary rules are sufficiently good enough.
You can crawl Wiktionary to get mapping for significant amount of words covered there.
Many languages which use hieroglyphs like Korean or Japanese have specialized software like Mecab https://sourceforge.net/projects/mecab to romanize their words. You can use Mecab to build a phonetic dictionary by converting words to romanized form and then simply applying rules to turn them into phones.
It is enough to transcribe few thousand most common words to bootstrap the dictionary.
Once dictionary is bootstrapped you can extend it to larger vocabulary with the g2p-seq2seq tool as described in previous chapter.