# Epitran
A library and tool for transliterating orthographic text as IPA (International Phonetic Alphabet).
## Usage
The Python modules ```epitran``` and ```epitran.vector``` can be used to easily write more sophisticated Python programs for deploying the **Epitran** mapping tables, preprocessors, and postprocessors. This is documented below.
If you wish to use Epitran to convert **English** to IPA, **you must install the Flite (including `lex_lookup`) as detailed below.**
## Using the `epitran` Module
### The Epitran class
The most general functionality in the `epitran` module is encapsulated in the very simple `Epitran` class:
Epitran(code, preproc=True, postproc=True, ligatures=False, cedict_file=None).
Its constructor takes one argument, `code`, the ISO 639-3 code of the language to be transliterated plus a hyphen plus a four letter code for the script (e.g. 'Latn' for Latin script, 'Cyrl' for Cyrillic script, and 'Arab' for a Perso-Arabic script). It also takes optional keyword arguments:
* `preproc` and `postproc` enable pre- and post-processors. These are enabled by default.
* `ligatures` enables non-standard IPA ligatures like "ʤ" and "ʨ".
* `cedict_file` gives the path to the [CC-CEDict](https://cc-cedict.org/wiki/) dictionary file (relevant only when working with Mandarin Chinese and which, because of licensing restrictions cannot be distributed with Epitran).
* `tones` allows IPA tones (˩˨˧˦˥) to be included and is needed for tonal languages like Vietnamese and Hokkien. By default, this value is false and will remove IPA tones from the transcription.
* For more options, type ```help(epitran.Epitran.__init__)``` into a Python terminal session
```python
>>> import epitran
>>> epi = epitran.Epitran('uig-Arab') # Uyghur in Perso-Arabic script
```
It is now possible to use the Epitran class for English and Mandarin Chinese (Simplified and Traditional) G2P as well as the other langugages that use Epitran's "classic" model. For Chinese, it is necessary to point the constructor to a copy of the [CC-CEDict](https://cc-cedict.org/wiki/) dictionary:
```python
>>> import epitran
>>> epi = epitran.Epitran('cmn-Hans', cedict_file='cedict_1_0_ts_utf-8_mdbg.txt')
```
The most useful public method of the Epitran class is `transliterate`:
Epitran.**transliterate**(text, normpunc=False, ligatures=False). Convert `text` (in Unicode-encoded orthography of the language specified in the constructor) to IPA, which is returned. `normpunc` enables punctuation normalization and `ligatures` enables non-standard IPA ligatures like "ʤ" and "ʨ". Usage is illustrated below (Python 2):
```python
>>> epi.transliterate(u'Düğün')
u'dy\u0270yn'
>>> print(epi.transliterate(u'Düğün'))
dyɰyn
```
Epitran.**word_to_tuples**(word, normpunc=False):
Takes a `word` (a Unicode string) in a supported orthography as input and returns a list of tuples with each tuple corresponding to an IPA segment of the word. The tuples have the following structure:
```
(
character_category :: String,
is_upper :: Integer,
orthographic_form :: Unicode String,
phonetic_form :: Unicode String,
segments :: List<Tuples>
)
```
Note that **word_to_tuples** is not implemented for all language-script pairs.
The codes for `character_category` are from the initial characters of the two character sequences listed in the "General Category" codes found in [Chapter 4 of the Unicode Standard](http://www.unicode.org/versions/Unicode8.0.0/ch04.pdf#G134153). For example, "L" corresponds to letters and "P" corresponds to production marks. The above data structure is likely to change in subsequent versions of the library. The structure of ```segments``` is as follows:
```
(
segment :: Unicode String,
vector :: List<Integer>
)
```
Here is an example of an interaction with ```word_to_tuples``` (Python 2):
```python
>>> import epitran
>>> epi = epitran.Epitran('tur-Latn')
>>> epi.word_to_tuples(u'Düğün')
[(u'L', 1, u'D', u'd', [(u'd', [-1, -1, 1, -1, -1, -1, -1, -1, 1, -1, -1, 1, 1, -1, -1, -1, -1, -1, -1, 0, -1])]), (u'L', 0, u'u\u0308', u'y', [(u'y', [1, 1, -1, 1, -1, -1, -1, 0, 1, -1, -1, -1, -1, -1, 1, 1, -1, -1, 1, 1, -1])]), (u'L', 0, u'g\u0306', u'\u0270', [(u'\u0270', [-1, 1, -1, 1, 0, -1, -1, 0, 1, -1, -1, 0, -1, 0, -1, 1, -1, 0, -1, 1, -1])]), (u'L', 0, u'u\u0308', u'y', [(u'y', [1, 1, -1, 1, -1, -1, -1, 0, 1, -1, -1, -1, -1, -1, 1, 1, -1, -1, 1, 1, -1])]), (u'L', 0, u'n', u'n', [(u'n', [-1, 1, 1, -1, -1, -1, 1, -1, 1, -1, -1, 1, 1, -1, -1, -1, -1, -1, -1, 0, -1])])]
```
### The Backoff class
Sometimes, when parsing text in more than one script, it is useful to employ a graceful backoff. If one language mode does not work, it can be useful to fall back to another, and so on. This functionality is provided by the Backoff class:
Backoff(lang_script_codes, cedict_file=None)
Note that the Backoff class does not currently support parameterized preprocessor and postprocessor application and does not support non-standard ligatures. It also does not support punctuation normalization. `lang_script_codes` is a list of codes like `eng-Latn` or `hin-Deva`. For example, if one was transcribing a Hindi text with many English loanwords and some stray characters of Simplified Chinese, one might use the following code (Python 3):
```python
from epitran.backoff import Backoff
>>> backoff = Backoff(['hin-Deva', 'eng-Latn', 'cmn-Hans'], cedict_file=‘cedict_1_0_ts_utf-8_mdbg.txt')
>>> backoff.transliterate('हिन्दी')
'ɦindiː'
>>> backoff.transliterate('English')
'ɪŋɡlɪʃ'
>>> backoff.transliterate('中文')
'ʈ͡ʂoŋwən'
```
Backoff works on a token-by-token basis: tokens that contain mixed scripts will be returned as the empty string, since they cannot be fully converted by any of the modes.
The Backoff class has the following public methods:
* **transliterate**: returns a unicode string of IPA phonemes
* **trans_list**: returns a list of IPA unicode strings, each of which is a
phoneme
* **xsampa_list**: returns a list of X-SAMPA (ASCII) strings, each of which is
phoneme
Consider the following example (Python 3):
```python
>>> backoff.transliterate('हिन्दी')
'ɦindiː'
>>> backoff.trans_list('हिन्दी')
['ɦ', 'i', 'n', 'd', 'iː']
>>> backoff.xsampa_list('हिन्दी')
['h\\', 'i', 'n', 'd', 'i:']
```
### DictFirst
The `DictFirst` class provides a simple alternative to the `Backoff` class. It
requires a dictionary of words known to be of Language A, one word per line in a
UTF-8 encoded text file. It accepts three arguments: the language-script code
for Language A, that for Language B, and a path to the dictionary file. It has one public method, `transliteration`, which works like `Epitran.transliterate` except that it returns the transliteration for Language A if the input token is in the dictionary; otherwise, it returns the Language B transliteration of the token:
```python
>>> import dictfirst
>>> df = dictfirst.DictFirst('tpi-Latn', 'eng-Latn', '../sample-dict.txt')
>>> df.transliterate('pela')
'pela'
>>> df.transliterate('pelo')
'pɛlow'
```
### Preprocessors, postprocessors, and their pitfalls
In order to build a maintainable orthography to phoneme mapper, it is sometimes necessary to employ preprocessors that make contextual substitutions of symbols before text is passed to a orthography-to-IPA mapping system that preserves relationships between input and output characters. This is particularly true of languages with a poor sound-symbols correspondence (like French and English). Languages like French are particularly good targets for this approach because the pronunciation of a given string of letters is highly predictable even though the individual symbols often do not map neatly into sounds. (Sound-symbol correspondence is so poor in English that effective English G2P systems rely heavily on pronouncing dictionaries.)
Preprocessing the inputs words to allow for straightforward grapheme-to-phoneme mappings (as is done in the current version of ```epitran``` for some languages) is advantageous because the restricted regular expression language used to write the preprocessing rules is more powerful than the language for the mapping rules and allows the equivalent of many mapping rules to be written with a single rule. Without them, providing ```epitran``` support for languages like French and German would not be practical. However, they do present some problems. Specifically, when using a language with a preprocessor, one **must** be aware that the input word will not always be identical to the concatenation of the orthographic strings (```orthographic_form```) output by ```Epitran.word_to_tuples```. Instead, the output of ```word_to_tuple``` will reflect the output of the preprocessor, which may delete, insert, and change letters in order to allow direct orthography-to-phoneme mapping at the next step. The same is true of other methods that rely on ```Epitran.word_to_tuple``` such as ```VectorsWithIPASpace.word_to_segs``` from the ```epitran.vector``` module.
For information on writing new pre- and post-processors, see the section on "[Extending Epitran with map files, preprocessors and postprocessors](#extending-epitran)", below.
## Using the ```epitran.vector``` Module
The ```epitran.vector``` module is also very simple. It contains one class, ```VectorsWithIPASpace```, including one method of interest, ```word_to_segs```:
The constructor for ```VectorsWithIPASpace``` takes two arguments:
- `code`: the language-script code for the language to be processed.
- `spaces`: the codes for the punctuation/symbol/IPA space in which the characters/segments from the data are expected to reside. The available spaces are listed [below](#language-support).
Its principle method is ```word_to_segs```:
VectorWithIPASpace.**word_to_segs**(word, normpunc=False). `word` is a Unicode string. If the keyword argument *normpunc* is set to True, punctuation discovered in `word` is normalized to ASCII equivalents.
A typical interaction with the ```VectorsWithIPASpace``` object via the ```word_to_segs``` method is illustrated here (Python 2):
```python
>>> import epitran.vector
>>> vwis = epitran.vector.VectorsWithIPASpace('uzb-Latn', ['uzb-Latn'])
>>> vwis.word_to_segs(u'darë')
[(u'L', 0, u'd', u'd\u032a', u'40', [-1, -1, 1, -1, -1, -1, -1, -1, 1, -1, -1, 1, 1, 1, -1, -1, -1, -1, -1, 0, -1]), (u'L', 0, u'a', u'a', u'37', [1, 1, -1, 1, -1, -1, -1, 0, 1, -1, -1, -1, -1, -1, -1, -1, 1, 1, -1, 1, -1]), (u'L', 0, u'r', u'r', u'54', [-1, 1, 1, 1, 0, -1, -1, -1, 1, -1, -1, 1, 1, -1, -1, 0, 0, 0, -1, 0, -1]), (u'L', 0, u'e\u0308', u'ja', u'46', [-1, 1, -1, 1, -1, -1, -1, 0, 1, -1, -1, -1, -1, 0, -1, 1, -1, -1, -1, 0, -1]), (u'L', 0, u'e\u0308', u'ja', u'37', [1, 1, -1, 1, -1, -1, -1, 0, 1, -1, -1, -1, -1, -1, -1, -1, 1, 1, -1, 1, -1])]
```
(It is important to note that, though the word that serves as input--*darë*--has four letters, the output contains five tuples because the last letter in *darë* actually corresponds to two IPA segments, /j/ and /a/.) The returned data structure is a list of tuples, each with the following structure:
```
(
character_category :: String,
is_upper :: Integer,
orthographic_form :: Unicode String,
phonetic_form :: Unicode String,
in_ipa_punc_space :: Integer,
phonological_feature_vector :: List<Integer>
)
```
A few notes are in order regarding this data structure:
- ```character_category``` is defined as part of the Unicode standard ([Chapter 4](http://www.unicode.org/versions/Unicode8.0.0/ch04.pdf#G134153)). It consists of a single, uppercase letter from the set {'L', 'M', 'N', 'P', 'S', 'Z', 'C'}.. The most frequent of these are 'L' (letter), 'N' (number), 'P' (punctuation), and 'Z' (separator [including separating white space]).
- ```is_upper``` consists only of integers from the set {0, 1}, with 0 indicating lowercase and 1 indicating uppercase.
- The integer in ```in_ipa_punc_space``` is an index to a list of known characters/segments such that, barring degenerate cases, each character or segment is assignmed a unique and globally consistant number. In cases where a character is encountered which is not in the known space, this field has the value -1.
- The length of the list ```phonological_feature_vector``` should be constant for any instantiation of the class (it is based on the number of features defined in panphon) but is--in principle--variable. The integers in this list are drawn from the set {-1, 0, 1}, with -1 corresponding to '-', 0 corresponding to '0', and 1 corresponding to '+'. For characters with no IPA equivalent, all values in the list are 0.
## <a name='language-support'></a>Language Support
### Transliteration Language/Script Pairs
| Code | Language (Script) |
|-------------|-------------------------|
| aar-Latn | Afar |
| aii-Syrc | Assyrian Neo-Aramaic |
| amh-Ethi | Amharic |
| amh-Ethi-pp | Amharic (more phonetic) |
| amh-Ethi-red| Amharic (reduced) |
| ara-Arab | Literary Arabic |
| ava-Cyrl | Avaric |
| aze-Cyrl | Azerbaijani (Cyrillic) |
| aze-Latn | Azerbaijani (Latin) |
| ben-Beng | Bengali |
| ben-Beng-red| Bengali (reduced) |
| bxk-Latn | Bukusu |
| cat-Latn | Catalan |
| ceb-Latn | Cebuano |
| ces-Latn | Czech |
| cjy-Latn | Jin (Wiktionary) |
| cmn-Hans | Mandarin (Simplified)\* |
| cmn-Hant | Mandarin (Traditional)\*|
| cmn-Latn | Mandarin (Pinyin)\* |
| ckb-Arab | Sorani |
| csb-Latn | Kashubian |
| deu-Latn | German |
| deu-Latn-np | German† |
| deu-Latn-nar| German (more phonetic) |
| eng-Latn | English‡ |
| fas-Arab | Farsi (Perso-Arabic) |
| fra-Latn | French |
| fra-Latn-np | French† |
| fra-Latn-p | French (more phonetic) |
| ful-Latn | Fulah |
| gan-Latn | Gan (Wiktionary) |
| got-Latn | Gothic |
| hak-Latn | Hakka (pha̍k-fa-sṳ) |
| hau-Latn | Hausa |
| hin-Deva | Hindi |
| hmn-Latn | Hmong |
| hrv-Latn | Croatian |
| hsn-Latn | Xiang (Wiktionary) |
| hun-Latn | Hungarian |
| ilo-Latn | Ilocano |
| ind-Latn | Indonesian |
| ita-Latn | Italian |
| jam-Latn | Jamaican |
| jav-Latn | Javanese |
| kaz-Cyrl | Kazakh (Cyrillic) |
| kaz-Cyrl-bab| Kazakh (Cyrillic—Babel) |
| kaz-Latn | Kazakh (Latin) |
| kbd-Cyrl | Kabardian |
| khm-Khmr | Khmer |
| kin-Latn | Kinyarwanda |
| kir-Arab | Kyrgyz (Perso-Arabic) |
| kir-Cyrl | Kyrgyz (Cyrillic) |
| kir-Latn | Kyrgyz (Latin) |
| kmr-Latn | Kurmanji |
| kmr-Latn-red| Kurmanji (reduced) |
| lao-Laoo | Lao |
| lij-Latn | Ligurian |
| lsm-Latn | Saamia |
| ltc-Latn-bax | Middle Chinese (Baxter and Sagart 2014) |
| mal-Mlym | Malayalam |
| mar-Deva | Marathi |
| mlt-Latn | Maltese |
| mon-Cyrl-bab| Mongolian (Cyrillic) |
| mri-Latn | Maori |
| msa-Latn | Malay |
| mya-Mymr | Burmese |
| nan-Latn | Hokkien (pe̍h-oē-jī) |
| nan-Latn-tl | Hokkien (Tâi-lô) |
| nld-Latn | Dutch |
| nya-Latn | Chichewa |
| ood-Lat-alv | Tohono O'odham |
| ood-Latn-sax| Tohono O'odham |
| ori-Orya | Odia |
| orm-Latn | Oromo |
| pan-Guru | Punjabi (Eastern) |
| pol-Latn | Polish |
| por-Latn | Portuguese |
| ron-Latn | Romanian |
| run-Latn | Rundi |
| rus-Cyrl | Russian |
| sag-Latn | Sango |
| sin-Sinh | Sinhala |
| sna-Latn | Shona |
| som-Latn | Somali |
| spa-Latn | Spanish |
| spa-Latn-eu | Spanish (Iberian) |
| sqi-Latn | Albanian |
| swa-Latn | Swahili |
| swa-Latn-red| Swahili (reduced) |
| swe-Latn | Swedish |
| tam-Taml | Tamil |
| tam-Taml-red| Tamil (reduced) |
| tel-Telu | Telugu |
| tgk-Cyrl | Tajik |
| tgl-Latn | Tagalog |
| tgl-Latn-red| Tagalog (reduced) |
| tha-Thai | Thai |
| tir-Ethi | Tigrinya |
| tir-Ethi-pp | Tigrinya (more phonemic)|
| tir-Ethi-red| Tigrinya (reduced) |
| tpi-Latn | Tok Pisin |
| tuk-Cyrl | Turkmen (Cyrillic) |
| tuk-Latn | Turkmen (Latin) |
| tur-Latn | Turkish (Latin) |
| tur-Latn-bab| Turkish (Latin—Babel) |
| tur-Latn-red| Turkish (reduced) |
| ukr-Cyrl | Ukranian |
| urd-Arab | Urdu |
| uig-Arab | Uyghur (Perso-Arabic) |
| uzb-Cyrl | Uzbek (Cyrillic) |
| uzb-Latn | Uzbek (Latin) |
| vie-Latn | Vietnamese |
| wuu-Latn | Shanghainese Wu (Wiktionary) |
| xho-Latn | Xhosa |
| yor-Latn | Yoruba |
| zha-Latn | Zhuang |
| zul-Latn | Zulu |
\*Chinese G2P requires the freely available [CC-CEDict](https://cc-cedict.org/wiki/) dictionary.
†These language preprocessors and maps naively assume a phonemic orthography.
‡English G2P requires the installation of the freely available [CMU Flite](http://tts.speech.cs.cmu.edu/awb/flite-2.0.5-current.tar.bz2) speech synthesis system.
### Languages with limited support due to highly ambiguous orthographies
Some the languages listed above should be approached with caution. It is not possible to provide highly accurate support for these language-script pairs due to the high degree of ambiguity inherent in the orthographies. Eventually, we plan to support these languages with a different back end based on WFSTs or neural methods.
| Code | Language (Script) |
|-------------|-------------------------|
| ara-Arab | Arabic |
| cat-Latn | Catalan |
| ckb-Arab | Sorani |
| fas-Arab | Farsi (Perso-Arabic) |
| fra-Latn | French |
| fra-Latn-np | French† |
| mya-Mymr | Burmese |
| por-Latn | Portuguese |
### Language "Spaces"
| Code | Language | Note |
|-----------------|----------|--------------------------------------|
| amh-Ethi | Amharic | |
| deu-Latn | German | |
| eng-Latn | English | |
| nld-Latn | Dutch | |
| spa-Latn | Spanish | |
| tur-Latn | Turkish | Based on data with suffixes attached |
| tur-Latn-nosuf | Turkish | Based on data with suffixes removed |
| uzb-Latn-suf | Uzbek | Based on data with suffixes attached |
Note that major languages, including **French**, are missing from this table due to a lack of appropriate text data.
## Installation of Flite (for English G2P)
For use with most languages, Epitran requires no special installation steps. It can be installed as an ordinarary python package, either with `pip` or by running `python setup.py install` in the root of the source directory. However, English G2P in Epitran relies on CMU Flite, a speech synthesis package by Alan Black and other speech researchers at Carnegie Mellon University. For the current version of Epitran, you should follow the installation instructions for `lex_lookup`, which is used as the default G2P interface for Epitran.
### `t2p`
**Not recommended** This interface to Flite is now deprecated; Use `lex_lookup`.
### `lex_lookup`
**Recommended**
`t2p` does not behave as expected on letter sequences that are highly infrequent in English. In such cases, `t2p` gives the pronunciation of the English letters of the name, rather than an attempt at the pronunciation of the name. There is a different binary included in the most recent (pre-release) versions of Flite that behaves better in this regard, but takes some extra effort to install. To install, you need to obtain at least version [2.0.5](http://tts.speech.cs.cmu.edu/awb/flite-2.0.5-current.tar.bz2) of Flite. We recommend that you obtain the source from GitHub (https://github.com/festvox/flite). Untar and compile the source, following the steps below, adjusting where appropriate for your system:
```bash
$ tar xjf flite-2.0.5-current.tar.bz2
$ cd flite-2.0.5-current
```
or
```bash
$ git clone git@github.com:festvox/flite.git
$ cd flite/
```
then
```bash
$ ./configure && make
$ sudo make install
$ cd testsuite
$ make lex_lookup
$ sudo cp lex_lookup /usr/local/bin
```
When installing on MacOS and other systems that use a BSD version of `cp`, some modification to a Makefile must be made in order to install flite-2.0.5 (between steps 3 and 4). Edit `main/Makefile` and change both instances of `cp -pd` to `cp -pR`. Then resume the steps above at step 4.
### Usage
To use `lex_lookup`, simply instantiate Epitran as usual, but with the `code` set to 'eng-Latn':
```python
>>> import epitran
>>> epi = epitran.Epitran('eng-Latn')
>>> print epi.transliterate(u'Berkeley')
bɹ̩kli
```
## <a name='extending-epitran'></a>Extending Epitran with map files, preprocessors and postprocessors
Language support in Epitran is provided through map files, which define mappings between orthographic and phonetic units, preprocessors that run before the map is applied, and postprocessors that run after the map is applied. Maps are defined in UTF8-encoded, comma-delimited value (CSV) files. The files are each named `<iso639>-<iso15924>.csv` where `<iso639>` is the (three letter, all lowercase) ISO 639-3 code for the language and `<iso15924>` is the (four letter, capitalized) ISO 15924 code for the script. These files reside in the `data` directory of the Epitran installation under the `map`, `pre`, and `post` subdirectories, respectively. The pre- and post-processor files are text files whose format is described belown. They follow the same naming conventions except that they have the file extensions `.txt`.
### Map files (mapping tables)
The map files are simple, two-column files where the first column contains the orthgraphic characters/sequences and the second column contains the phonetic characters/sequences. The two columns are separated by a comma; each row is terminated by a newline. For many languages (most languages with unambiguous, phonemically adequate orthographies) just this easy-to-produce mapping file is adequate to produce a serviceable G2P system.
The first row is a header and is discarded. For consistency, it should contain the fields "Orth" and "Phon". The following rows by consist of fields of any length, separated by a comma. The same phonetic form (the second field) may occur any number of times but an orthographic form may only occur once. Where one orthograrphic form is a prefix of another form, the longer form has priority in mapping. In other words, matching between orthographic units and orthographic strings is greedy. Mapping works by finding the longest prefix of the orthographic form and adding the corresponding phonetic string to the end of the phonetic form, then removing the prefix from the orthographic form and continuing, in the same manner, until the orthographic form is consumed. If no non-empty prefix of the orthographic form is present in the mapping table, the first character in the orthographic form is removed and appended to the phonetic form. The normal sequence then resumes. This means that non-phonetic characters may end up in the "phonetic" form, which we judge to be better than losing information through an inadequate mapping table.
### Preprocesssors and postprocessors
For language-script pairs with more complicated orthographies, it is sometimes necessary to manipulate the orthographic form prior to mapping or to manipulate the phonetic form after mapping. This is done, in Epitran, with grammars of context-sensitive string rewrite rules. In truth, these rules would be more than adequate to solve the mapping problem as well but in practical terms, it is usually easier to let easy-to-understand and easy-to-maintain mapping files carry most of the weight of conversion and reserve the more powerful context sensitive grammar formalism for pre- and post-processing.
The preprocessor and postprocessor files have the same format. They consist of a sequence of lines, each consisting of one of four types:
1. Symbol definitions
2. Context-sensitive rewrite rules
3. Comments
4. Blank lines
#### Symbol definitions
Lines like the following
```
::vowels:: = a|e|i|o|u
```
define symbols that can be reused in writing rules. Symbols must consist of a prefix of two colons, a sequence of one or more lowercase letters and underscores, and a suffix of two colons. The are separated from their definitions by the equals sign (optionally set off with white space). The definition consists of a substring from a regular expression.
Symbols must be defined before they are referenced.
#### Rewrite rules
Context-sensitive rewrite rules in Epitran are written in a format familiar to phonologists but transparent to computer scientists. They can be schematized as
```
a -> b / X _ Y
```
which can be rewitten as
```
XaY → XbY
```
The arrow `->` can be read as "is rewritten as" and the slash `/` can be read as "in the context". The underscore indicates the position of the symbol(s) being rewritten. Another special symbol is the octothorp `#`, which indicates the beginning or end of a (word length) string (a word boundary). Consider the following rule:
```
e -> ə / _ #
```
This rule can be read as "/e/ is rewritten as /ə/ in the context at the end of the word." A final special symbol is zero `0`, which represents the empty string. It is used in rules that insert or delete segments. Consider the following rule that deletes /ə/ between /k/ and /l/:
```
ə -> 0 / k _ l
```
All rules must include the arrow operator, the slash operator, and the underscore. A rule that applies in a context-free fashion can be written in the following way:
```
ch -> x / _
```
The implementation of context-sensitive rules in Epitran pre- and post-processors uses regular expression replacement. Specifically, it employs the `regex` package, a drop-in replacement for `re`. Because of this, regular expression notation can be used in writing rules:
```
c -> s / _ [ie]
```
or
```
c -> s / _ (i|e)
```
For a complete guide to `regex` regular expressions, see the documentation for [`re`](https://docs.python.org/2/library/re.html) and for [`regex`](https://pypi.python.org/pypi/regex), specifically.
Fragments of regular expressions can be assigned to symbols and reused throughout a file. For example, symbol for the disjunction of vowels in a language can be used in a rule that changes /u/ into /w/ before vowels:
```
::vowels:: = a|e|i|o|u
...
u -> w / _ (::vowels::)
```
There is a special construct for handling cases of metathesis (where "AB" is replaced with "BA"). For example, the rule:
```
(?P<sw1>[เแโไใไ])(?P<sw2>.) -> 0 / _
```
Will "swap" the positions of any character in "เแโไใไ" and any following character. Left of the arrow, there should be two groups (surrounded by parentheses) with the names `sw1` and `sw2` (a name for a group is specified by `?P<name>` appearing immediately after the open parenthesis for a group). The substrings matched by the two groups, `sw1` and `sw2` will be "swapped" or metathesized. The item immediately right of the arrow is ignored, but the context is not.
To move IPA tones to the end of the word, first ensure that tones=True in the instantiated Epitran object and use the following rule:
```
(?P<sw1>[˩˨˧˦˥]+)(?P<sw2>\w+) -> 0 / _\b
```
The rules apply in order, so earlier rules may "feed" and "bleed" later rules. Therefore, their sequence is *very important* and can be leveraged in order to achieve valuable results.
#### Comments and blank lines
Comments and blank lines (lines consisting only of white space) are allowed to make your code more readable. Any line in which the first non-whitespace character is a percent sign `%` is interpreted as comment. The rest of the line is ignored when the file is interpreted. Blank lines are also ignored.
### A strategy for adding language support
Epitran uses a mapping-and-repairs approach to G2P. It is expected that there is a mapping between graphemes and phonemes that can do most of the work of converting orthographic representations to phonological representations. In phonemically adequate orthogrphies, this mapping can do *all* of the work. This mapping should be completed first. For many languages, a basis for this mapping table already exists on [Wikipedia](http://www.wikipedia.org) and [Omniglot](http://www.omniglot.com) (though the Omniglot tables are typically not machine readable).
On the other hand, many writing systems deviate from the phonemically adequate idea. It is here that pre- and post-processors must be introduced. For example, in Swedish, the letter `<a>` receives a different pronunciation before two consonants (/ɐ/) than elsewhere (/ɑː/). It makes sense to add a preprocessor rule that rewrites `<a>` as /ɐ/ before two consonants (and similar rules for the other vowels, since they are affected by the same condition). Preprocessor rules should generally be employed whenever the orthographic representation must be adjusted (by contextual changes, deletions, etc.) prior to the mapping step.
One common use for postprocessors is to eliminate characters that are needed by the preprocessors or maps, but which should not appear in the output. A classic example of this is the virama used in Indic scripts. In these scripts, in order to write a consonant *not followed* by a vowel, one uses the form of the consonant symbol with particular inherent vowel followed by a virama (which has various names in different Indic languages). An easy way of handling this is to allow the mapping to translate the consonant into an IPA consonant + an inherent vowel (which, for a given language, will always be the same), then use the postprocessor to delete the vowel + virama sequence (wherever it occurs).
In fact, any situation where a character that is introduced by the map needs to be subsequently deleted is a good use-case for postprocessors. Another example from Indian languages includes so-called schwa deletion. Some vowels implied by a direct mapping between the orthography and the phonology are not actually pronounced; these vowels can generally be predicted. In most languages, they occur in the context after a vowel+consonant sequence and before a consonant+vowel sequence. In other words, the rule looks like the following:
```
ə -> 0 / (::vowel::)(::consonant::) _ (::consonant::)(::vowel::)
```
Perhaps the best way to learn how to structure language support for a new language is to consult the existing languages in Epitran. The French preprocessor `fra-Latn.txt` and the Thai postprocessor `tha-Thai.txt` illustrate many of the use-cases for these rules.
# Citing Epitran
If you use Epitran in published work, or in other research, please use the following citation:
David R. Mortensen, Siddharth Dalmia, and Patrick Littell. 2018. Epitran: Precision G2P for many languages. In *Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)*, Paris, France. European Language Resources Association (ELRA).
```
@InProceedings{Mortensen-et-al:2018,
author = {Mortensen, David R. and Dalmia, Siddharth and Littell, Patrick},
title = {Epitran: Precision {G2P} for Many Languages},
booktitle = {Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)},
year = {2018},
month = {May},
date = {7--12},
location = {Miyazaki, Japan},
editor = {Nicoletta Calzolari (Conference chair) and Khalid Choukri and Christopher Cieri and Thierry Declerck and Sara Goggi and Koiti Hasida and Hitoshi Isahara and Bente Maegaard and Joseph Mariani and H\'el\`ene Mazo and Asuncion Moreno and Jan Odijk and Stelios Piperidis and Takenobu Tokunaga},
publisher = {European Language Resources Association (ELRA)},
address = {Paris, France},
isbn = {979-10-95546-00-9},
language = {english}
}
```
Raw data
{
"_id": null,
"home_page": "http://github.com/dmort27/epitran",
"name": "epitran",
"maintainer": "",
"docs_url": null,
"requires_python": "",
"maintainer_email": "",
"keywords": "",
"author": "David R. Mortensen",
"author_email": "dmortens@cs.cmu.edu",
"download_url": "https://files.pythonhosted.org/packages/ed/85/e2a6b5ae4c734d30c25fd74dcbdc76af68b524de46c9659e8eb9bcd511e4/epitran-1.25.tar.gz",
"platform": null,
"description": "# Epitran\n\nA library and tool for transliterating orthographic text as IPA (International Phonetic Alphabet).\n\n## Usage\n\nThe Python modules ```epitran``` and ```epitran.vector``` can be used to easily write more sophisticated Python programs for deploying the **Epitran** mapping tables, preprocessors, and postprocessors. This is documented below.\n\nIf you wish to use Epitran to convert **English** to IPA, **you must install the Flite (including `lex_lookup`) as detailed below.**\n\n## Using the `epitran` Module\n\n### The Epitran class\n\nThe most general functionality in the `epitran` module is encapsulated in the very simple `Epitran` class:\n\nEpitran(code, preproc=True, postproc=True, ligatures=False, cedict_file=None).\n\nIts constructor takes one argument, `code`, the ISO 639-3 code of the language to be transliterated plus a hyphen plus a four letter code for the script (e.g. 'Latn' for Latin script, 'Cyrl' for Cyrillic script, and 'Arab' for a Perso-Arabic script). It also takes optional keyword arguments:\n* `preproc` and `postproc` enable pre- and post-processors. These are enabled by default.\n* `ligatures` enables non-standard IPA ligatures like \"\u02a4\" and \"\u02a8\".\n* `cedict_file` gives the path to the [CC-CEDict](https://cc-cedict.org/wiki/) dictionary file (relevant only when working with Mandarin Chinese and which, because of licensing restrictions cannot be distributed with Epitran).\n* `tones` allows IPA tones (\u02e9\u02e8\u02e7\u02e6\u02e5) to be included and is needed for tonal languages like Vietnamese and Hokkien. By default, this value is false and will remove IPA tones from the transcription. \n* For more options, type ```help(epitran.Epitran.__init__)``` into a Python terminal session\n\n```python\n>>> import epitran\n>>> epi = epitran.Epitran('uig-Arab') # Uyghur in Perso-Arabic script\n```\n\nIt is now possible to use the Epitran class for English and Mandarin Chinese (Simplified and Traditional) G2P as well as the other langugages that use Epitran's \"classic\" model. For Chinese, it is necessary to point the constructor to a copy of the [CC-CEDict](https://cc-cedict.org/wiki/) dictionary:\n\n\n```python\n>>> import epitran\n>>> epi = epitran.Epitran('cmn-Hans', cedict_file='cedict_1_0_ts_utf-8_mdbg.txt')\n```\n\nThe most useful public method of the Epitran class is `transliterate`:\n\nEpitran.**transliterate**(text, normpunc=False, ligatures=False). Convert `text` (in Unicode-encoded orthography of the language specified in the constructor) to IPA, which is returned. `normpunc` enables punctuation normalization and `ligatures` enables non-standard IPA ligatures like \"\u02a4\" and \"\u02a8\". Usage is illustrated below (Python 2):\n\n\n```python\n>>> epi.transliterate(u'D\u00fc\u011f\u00fcn')\nu'dy\\u0270yn'\n>>> print(epi.transliterate(u'D\u00fc\u011f\u00fcn'))\ndy\u0270yn\n```\n\n\nEpitran.**word_to_tuples**(word, normpunc=False):\nTakes a `word` (a Unicode string) in a supported orthography as input and returns a list of tuples with each tuple corresponding to an IPA segment of the word. The tuples have the following structure:\n```\n(\n character_category :: String,\n is_upper :: Integer,\n orthographic_form :: Unicode String,\n phonetic_form :: Unicode String,\n segments :: List<Tuples>\n)\n```\n\nNote that **word_to_tuples** is not implemented for all language-script pairs.\n\nThe codes for `character_category` are from the initial characters of the two character sequences listed in the \"General Category\" codes found in [Chapter 4 of the Unicode Standard](http://www.unicode.org/versions/Unicode8.0.0/ch04.pdf#G134153). For example, \"L\" corresponds to letters and \"P\" corresponds to production marks. The above data structure is likely to change in subsequent versions of the library. The structure of ```segments``` is as follows:\n```\n(\n segment :: Unicode String,\n vector :: List<Integer>\n)\n```\n\nHere is an example of an interaction with ```word_to_tuples``` (Python 2):\n\n\n```python\n>>> import epitran\n>>> epi = epitran.Epitran('tur-Latn')\n>>> epi.word_to_tuples(u'D\u00fc\u011f\u00fcn')\n[(u'L', 1, u'D', u'd', [(u'd', [-1, -1, 1, -1, -1, -1, -1, -1, 1, -1, -1, 1, 1, -1, -1, -1, -1, -1, -1, 0, -1])]), (u'L', 0, u'u\\u0308', u'y', [(u'y', [1, 1, -1, 1, -1, -1, -1, 0, 1, -1, -1, -1, -1, -1, 1, 1, -1, -1, 1, 1, -1])]), (u'L', 0, u'g\\u0306', u'\\u0270', [(u'\\u0270', [-1, 1, -1, 1, 0, -1, -1, 0, 1, -1, -1, 0, -1, 0, -1, 1, -1, 0, -1, 1, -1])]), (u'L', 0, u'u\\u0308', u'y', [(u'y', [1, 1, -1, 1, -1, -1, -1, 0, 1, -1, -1, -1, -1, -1, 1, 1, -1, -1, 1, 1, -1])]), (u'L', 0, u'n', u'n', [(u'n', [-1, 1, 1, -1, -1, -1, 1, -1, 1, -1, -1, 1, 1, -1, -1, -1, -1, -1, -1, 0, -1])])]\n```\n\n### The Backoff class\n\nSometimes, when parsing text in more than one script, it is useful to employ a graceful backoff. If one language mode does not work, it can be useful to fall back to another, and so on. This functionality is provided by the Backoff class:\n\nBackoff(lang_script_codes, cedict_file=None)\n\nNote that the Backoff class does not currently support parameterized preprocessor and postprocessor application and does not support non-standard ligatures. It also does not support punctuation normalization. `lang_script_codes` is a list of codes like `eng-Latn` or `hin-Deva`. For example, if one was transcribing a Hindi text with many English loanwords and some stray characters of Simplified Chinese, one might use the following code (Python 3):\n\n```python\nfrom epitran.backoff import Backoff\n>>> backoff = Backoff(['hin-Deva', 'eng-Latn', 'cmn-Hans'], cedict_file=\u2018cedict_1_0_ts_utf-8_mdbg.txt')\n>>> backoff.transliterate('\u0939\u093f\u0928\u094d\u0926\u0940')\n'\u0266indi\u02d0'\n>>> backoff.transliterate('English')\n'\u026a\u014b\u0261l\u026a\u0283'\n>>> backoff.transliterate('\u4e2d\u6587')\n'\u0288\u0361\u0282o\u014bw\u0259n'\n```\n\nBackoff works on a token-by-token basis: tokens that contain mixed scripts will be returned as the empty string, since they cannot be fully converted by any of the modes.\n\nThe Backoff class has the following public methods:\n* **transliterate**: returns a unicode string of IPA phonemes\n* **trans_list**: returns a list of IPA unicode strings, each of which is a\n phoneme\n* **xsampa_list**: returns a list of X-SAMPA (ASCII) strings, each of which is\n phoneme\n\nConsider the following example (Python 3):\n\n```python\n>>> backoff.transliterate('\u0939\u093f\u0928\u094d\u0926\u0940')\n'\u0266indi\u02d0'\n>>> backoff.trans_list('\u0939\u093f\u0928\u094d\u0926\u0940')\n['\u0266', 'i', 'n', 'd', 'i\u02d0']\n>>> backoff.xsampa_list('\u0939\u093f\u0928\u094d\u0926\u0940')\n['h\\\\', 'i', 'n', 'd', 'i:']\n```\n\n### DictFirst\n\nThe `DictFirst` class provides a simple alternative to the `Backoff` class. It\nrequires a dictionary of words known to be of Language A, one word per line in a\nUTF-8 encoded text file. It accepts three arguments: the language-script code\nfor Language A, that for Language B, and a path to the dictionary file. It has one public method, `transliteration`, which works like `Epitran.transliterate` except that it returns the transliteration for Language A if the input token is in the dictionary; otherwise, it returns the Language B transliteration of the token:\n\n```python\n>>> import dictfirst\n>>> df = dictfirst.DictFirst('tpi-Latn', 'eng-Latn', '../sample-dict.txt')\n>>> df.transliterate('pela')\n'pela'\n>>> df.transliterate('pelo')\n'p\u025blow'\n```\n\n### Preprocessors, postprocessors, and their pitfalls\n\nIn order to build a maintainable orthography to phoneme mapper, it is sometimes necessary to employ preprocessors that make contextual substitutions of symbols before text is passed to a orthography-to-IPA mapping system that preserves relationships between input and output characters. This is particularly true of languages with a poor sound-symbols correspondence (like French and English). Languages like French are particularly good targets for this approach because the pronunciation of a given string of letters is highly predictable even though the individual symbols often do not map neatly into sounds. (Sound-symbol correspondence is so poor in English that effective English G2P systems rely heavily on pronouncing dictionaries.)\n\nPreprocessing the inputs words to allow for straightforward grapheme-to-phoneme mappings (as is done in the current version of ```epitran``` for some languages) is advantageous because the restricted regular expression language used to write the preprocessing rules is more powerful than the language for the mapping rules and allows the equivalent of many mapping rules to be written with a single rule. Without them, providing ```epitran``` support for languages like French and German would not be practical. However, they do present some problems. Specifically, when using a language with a preprocessor, one **must** be aware that the input word will not always be identical to the concatenation of the orthographic strings (```orthographic_form```) output by ```Epitran.word_to_tuples```. Instead, the output of ```word_to_tuple``` will reflect the output of the preprocessor, which may delete, insert, and change letters in order to allow direct orthography-to-phoneme mapping at the next step. The same is true of other methods that rely on ```Epitran.word_to_tuple``` such as ```VectorsWithIPASpace.word_to_segs``` from the ```epitran.vector``` module.\n\nFor information on writing new pre- and post-processors, see the section on \"[Extending Epitran with map files, preprocessors and postprocessors](#extending-epitran)\", below.\n\n## Using the ```epitran.vector``` Module\n\nThe ```epitran.vector``` module is also very simple. It contains one class, ```VectorsWithIPASpace```, including one method of interest, ```word_to_segs```:\n\nThe constructor for ```VectorsWithIPASpace``` takes two arguments:\n- `code`: the language-script code for the language to be processed.\n- `spaces`: the codes for the punctuation/symbol/IPA space in which the characters/segments from the data are expected to reside. The available spaces are listed [below](#language-support).\n\nIts principle method is ```word_to_segs```:\n\nVectorWithIPASpace.**word_to_segs**(word, normpunc=False). `word` is a Unicode string. If the keyword argument *normpunc* is set to True, punctuation discovered in `word` is normalized to ASCII equivalents.\n\n\nA typical interaction with the ```VectorsWithIPASpace``` object via the ```word_to_segs``` method is illustrated here (Python 2):\n\n```python\n>>> import epitran.vector\n>>> vwis = epitran.vector.VectorsWithIPASpace('uzb-Latn', ['uzb-Latn'])\n>>> vwis.word_to_segs(u'dar\u00eb')\n[(u'L', 0, u'd', u'd\\u032a', u'40', [-1, -1, 1, -1, -1, -1, -1, -1, 1, -1, -1, 1, 1, 1, -1, -1, -1, -1, -1, 0, -1]), (u'L', 0, u'a', u'a', u'37', [1, 1, -1, 1, -1, -1, -1, 0, 1, -1, -1, -1, -1, -1, -1, -1, 1, 1, -1, 1, -1]), (u'L', 0, u'r', u'r', u'54', [-1, 1, 1, 1, 0, -1, -1, -1, 1, -1, -1, 1, 1, -1, -1, 0, 0, 0, -1, 0, -1]), (u'L', 0, u'e\\u0308', u'ja', u'46', [-1, 1, -1, 1, -1, -1, -1, 0, 1, -1, -1, -1, -1, 0, -1, 1, -1, -1, -1, 0, -1]), (u'L', 0, u'e\\u0308', u'ja', u'37', [1, 1, -1, 1, -1, -1, -1, 0, 1, -1, -1, -1, -1, -1, -1, -1, 1, 1, -1, 1, -1])]\n```\n\n\n(It is important to note that, though the word that serves as input--*dar\u00eb*--has four letters, the output contains five tuples because the last letter in *dar\u00eb* actually corresponds to two IPA segments, /j/ and /a/.) The returned data structure is a list of tuples, each with the following structure:\n\n```\n(\n character_category :: String,\n is_upper :: Integer,\n orthographic_form :: Unicode String,\n phonetic_form :: Unicode String,\n in_ipa_punc_space :: Integer,\n phonological_feature_vector :: List<Integer>\n)\n```\n\nA few notes are in order regarding this data structure:\n- ```character_category``` is defined as part of the Unicode standard ([Chapter 4](http://www.unicode.org/versions/Unicode8.0.0/ch04.pdf#G134153)). It consists of a single, uppercase letter from the set {'L', 'M', 'N', 'P', 'S', 'Z', 'C'}.. The most frequent of these are 'L' (letter), 'N' (number), 'P' (punctuation), and 'Z' (separator [including separating white space]).\n- ```is_upper``` consists only of integers from the set {0, 1}, with 0 indicating lowercase and 1 indicating uppercase.\n- The integer in ```in_ipa_punc_space``` is an index to a list of known characters/segments such that, barring degenerate cases, each character or segment is assignmed a unique and globally consistant number. In cases where a character is encountered which is not in the known space, this field has the value -1.\n- The length of the list ```phonological_feature_vector``` should be constant for any instantiation of the class (it is based on the number of features defined in panphon) but is--in principle--variable. The integers in this list are drawn from the set {-1, 0, 1}, with -1 corresponding to '-', 0 corresponding to '0', and 1 corresponding to '+'. For characters with no IPA equivalent, all values in the list are 0.\n\n\n## <a name='language-support'></a>Language Support\n\n### Transliteration Language/Script Pairs\n\n| Code | Language (Script) |\n|-------------|-------------------------|\n| aar-Latn | Afar |\n| aii-Syrc | Assyrian Neo-Aramaic |\n| amh-Ethi | Amharic |\n| amh-Ethi-pp | Amharic (more phonetic) |\n| amh-Ethi-red| Amharic (reduced) |\n| ara-Arab | Literary Arabic |\n| ava-Cyrl | Avaric |\n| aze-Cyrl | Azerbaijani (Cyrillic) |\n| aze-Latn | Azerbaijani (Latin) |\n| ben-Beng | Bengali |\n| ben-Beng-red| Bengali (reduced) |\n| bxk-Latn | Bukusu |\n| cat-Latn | Catalan |\n| ceb-Latn | Cebuano |\n| ces-Latn | Czech |\n| cjy-Latn | Jin (Wiktionary) |\n| cmn-Hans | Mandarin (Simplified)\\* |\n| cmn-Hant | Mandarin (Traditional)\\*|\n| cmn-Latn | Mandarin (Pinyin)\\* |\n| ckb-Arab | Sorani |\n| csb-Latn | Kashubian |\n| deu-Latn | German |\n| deu-Latn-np | German\u2020 |\n| deu-Latn-nar| German (more phonetic) |\n| eng-Latn | English\u2021 |\n| fas-Arab | Farsi (Perso-Arabic) |\n| fra-Latn | French |\n| fra-Latn-np | French\u2020 |\n| fra-Latn-p | French (more phonetic) |\n| ful-Latn | Fulah |\n| gan-Latn | Gan (Wiktionary) |\n| got-Latn | Gothic |\n| hak-Latn | Hakka (pha\u030dk-fa-s\u1e73) |\n| hau-Latn | Hausa |\n| hin-Deva | Hindi |\n| hmn-Latn | Hmong |\n| hrv-Latn | Croatian |\n| hsn-Latn | Xiang (Wiktionary) |\n| hun-Latn | Hungarian |\n| ilo-Latn | Ilocano |\n| ind-Latn | Indonesian |\n| ita-Latn | Italian |\n| jam-Latn | Jamaican |\n| jav-Latn | Javanese |\n| kaz-Cyrl | Kazakh (Cyrillic) |\n| kaz-Cyrl-bab| Kazakh (Cyrillic\u2014Babel) |\n| kaz-Latn | Kazakh (Latin) |\n| kbd-Cyrl | Kabardian |\n| khm-Khmr | Khmer |\n| kin-Latn | Kinyarwanda |\n| kir-Arab | Kyrgyz (Perso-Arabic) |\n| kir-Cyrl | Kyrgyz (Cyrillic) |\n| kir-Latn | Kyrgyz (Latin) |\n| kmr-Latn | Kurmanji |\n| kmr-Latn-red| Kurmanji (reduced) |\n| lao-Laoo | Lao |\n| lij-Latn | Ligurian |\n| lsm-Latn | Saamia |\n| ltc-Latn-bax | Middle Chinese (Baxter and Sagart 2014) |\n| mal-Mlym | Malayalam |\n| mar-Deva | Marathi |\n| mlt-Latn | Maltese |\n| mon-Cyrl-bab| Mongolian (Cyrillic) |\n| mri-Latn | Maori |\n| msa-Latn | Malay |\n| mya-Mymr | Burmese |\n| nan-Latn | Hokkien (pe\u030dh-o\u0113-j\u012b) |\n| nan-Latn-tl | Hokkien (T\u00e2i-l\u00f4) |\n| nld-Latn | Dutch |\n| nya-Latn | Chichewa |\n| ood-Lat-alv | Tohono O'odham |\n| ood-Latn-sax| Tohono O'odham |\n| ori-Orya | Odia |\n| orm-Latn | Oromo |\n| pan-Guru | Punjabi (Eastern) |\n| pol-Latn | Polish |\n| por-Latn | Portuguese |\n| ron-Latn | Romanian |\n| run-Latn | Rundi |\n| rus-Cyrl | Russian |\n| sag-Latn | Sango |\n| sin-Sinh | Sinhala |\n| sna-Latn | Shona |\n| som-Latn | Somali |\n| spa-Latn | Spanish |\n| spa-Latn-eu | Spanish (Iberian) |\n| sqi-Latn | Albanian |\n| swa-Latn | Swahili |\n| swa-Latn-red| Swahili (reduced) |\n| swe-Latn | Swedish |\n| tam-Taml | Tamil |\n| tam-Taml-red| Tamil (reduced) |\n| tel-Telu | Telugu |\n| tgk-Cyrl | Tajik |\n| tgl-Latn | Tagalog |\n| tgl-Latn-red| Tagalog (reduced) |\n| tha-Thai | Thai |\n| tir-Ethi | Tigrinya |\n| tir-Ethi-pp | Tigrinya (more phonemic)|\n| tir-Ethi-red| Tigrinya (reduced) |\n| tpi-Latn | Tok Pisin |\n| tuk-Cyrl | Turkmen (Cyrillic) |\n| tuk-Latn | Turkmen (Latin) |\n| tur-Latn | Turkish (Latin) |\n| tur-Latn-bab| Turkish (Latin\u2014Babel) |\n| tur-Latn-red| Turkish (reduced) |\n| ukr-Cyrl | Ukranian |\n| urd-Arab | Urdu |\n| uig-Arab | Uyghur (Perso-Arabic) |\n| uzb-Cyrl | Uzbek (Cyrillic) |\n| uzb-Latn | Uzbek (Latin) |\n| vie-Latn | Vietnamese |\n| wuu-Latn | Shanghainese Wu (Wiktionary) |\n| xho-Latn | Xhosa |\n| yor-Latn | Yoruba |\n| zha-Latn | Zhuang |\n| zul-Latn | Zulu |\n\n\\*Chinese G2P requires the freely available [CC-CEDict](https://cc-cedict.org/wiki/) dictionary.\n\n\u2020These language preprocessors and maps naively assume a phonemic orthography.\n\n\u2021English G2P requires the installation of the freely available [CMU Flite](http://tts.speech.cs.cmu.edu/awb/flite-2.0.5-current.tar.bz2) speech synthesis system.\n\n### Languages with limited support due to highly ambiguous orthographies\n\nSome the languages listed above should be approached with caution. It is not possible to provide highly accurate support for these language-script pairs due to the high degree of ambiguity inherent in the orthographies. Eventually, we plan to support these languages with a different back end based on WFSTs or neural methods.\n\n| Code | Language (Script) |\n|-------------|-------------------------|\n| ara-Arab | Arabic |\n| cat-Latn | Catalan |\n| ckb-Arab | Sorani |\n| fas-Arab | Farsi (Perso-Arabic) |\n| fra-Latn | French |\n| fra-Latn-np | French\u2020 |\n| mya-Mymr | Burmese |\n| por-Latn | Portuguese |\n\n### Language \"Spaces\"\n\n| Code | Language | Note |\n|-----------------|----------|--------------------------------------|\n| amh-Ethi | Amharic | |\n| deu-Latn | German | |\n| eng-Latn | English | |\n| nld-Latn | Dutch | |\n| spa-Latn | Spanish | |\n| tur-Latn | Turkish | Based on data with suffixes attached |\n| tur-Latn-nosuf | Turkish | Based on data with suffixes removed |\n| uzb-Latn-suf | Uzbek | Based on data with suffixes attached |\n\nNote that major languages, including **French**, are missing from this table due to a lack of appropriate text data.\n\n## Installation of Flite (for English G2P)\n\nFor use with most languages, Epitran requires no special installation steps. It can be installed as an ordinarary python package, either with `pip` or by running `python setup.py install` in the root of the source directory. However, English G2P in Epitran relies on CMU Flite, a speech synthesis package by Alan Black and other speech researchers at Carnegie Mellon University. For the current version of Epitran, you should follow the installation instructions for `lex_lookup`, which is used as the default G2P interface for Epitran.\n\n### `t2p`\n\n**Not recommended** This interface to Flite is now deprecated; Use `lex_lookup`.\n\n### `lex_lookup`\n\n**Recommended**\n\n`t2p` does not behave as expected on letter sequences that are highly infrequent in English. In such cases, `t2p` gives the pronunciation of the English letters of the name, rather than an attempt at the pronunciation of the name. There is a different binary included in the most recent (pre-release) versions of Flite that behaves better in this regard, but takes some extra effort to install. To install, you need to obtain at least version [2.0.5](http://tts.speech.cs.cmu.edu/awb/flite-2.0.5-current.tar.bz2) of Flite. We recommend that you obtain the source from GitHub (https://github.com/festvox/flite). Untar and compile the source, following the steps below, adjusting where appropriate for your system:\n\n\n```bash\n$ tar xjf flite-2.0.5-current.tar.bz2\n$ cd flite-2.0.5-current\n```\n\nor\n\n```bash\n$ git clone git@github.com:festvox/flite.git\n$ cd flite/\n```\n\nthen\n\n```bash\n$ ./configure && make\n$ sudo make install\n$ cd testsuite\n$ make lex_lookup\n$ sudo cp lex_lookup /usr/local/bin\n```\n\n\nWhen installing on MacOS and other systems that use a BSD version of `cp`, some modification to a Makefile must be made in order to install flite-2.0.5 (between steps 3 and 4). Edit `main/Makefile` and change both instances of `cp -pd` to `cp -pR`. Then resume the steps above at step 4.\n\n### Usage\n\nTo use `lex_lookup`, simply instantiate Epitran as usual, but with the `code` set to 'eng-Latn':\n\n\n```python\n>>> import epitran\n>>> epi = epitran.Epitran('eng-Latn')\n>>> print epi.transliterate(u'Berkeley')\nb\u0279\u0329kli\n```\n\n\n## <a name='extending-epitran'></a>Extending Epitran with map files, preprocessors and postprocessors\n\nLanguage support in Epitran is provided through map files, which define mappings between orthographic and phonetic units, preprocessors that run before the map is applied, and postprocessors that run after the map is applied. Maps are defined in UTF8-encoded, comma-delimited value (CSV) files. The files are each named `<iso639>-<iso15924>.csv` where `<iso639>` is the (three letter, all lowercase) ISO 639-3 code for the language and `<iso15924>` is the (four letter, capitalized) ISO 15924 code for the script. These files reside in the `data` directory of the Epitran installation under the `map`, `pre`, and `post` subdirectories, respectively. The pre- and post-processor files are text files whose format is described belown. They follow the same naming conventions except that they have the file extensions `.txt`.\n\n### Map files (mapping tables)\n\nThe map files are simple, two-column files where the first column contains the orthgraphic characters/sequences and the second column contains the phonetic characters/sequences. The two columns are separated by a comma; each row is terminated by a newline. For many languages (most languages with unambiguous, phonemically adequate orthographies) just this easy-to-produce mapping file is adequate to produce a serviceable G2P system.\n\nThe first row is a header and is discarded. For consistency, it should contain the fields \"Orth\" and \"Phon\". The following rows by consist of fields of any length, separated by a comma. The same phonetic form (the second field) may occur any number of times but an orthographic form may only occur once. Where one orthograrphic form is a prefix of another form, the longer form has priority in mapping. In other words, matching between orthographic units and orthographic strings is greedy. Mapping works by finding the longest prefix of the orthographic form and adding the corresponding phonetic string to the end of the phonetic form, then removing the prefix from the orthographic form and continuing, in the same manner, until the orthographic form is consumed. If no non-empty prefix of the orthographic form is present in the mapping table, the first character in the orthographic form is removed and appended to the phonetic form. The normal sequence then resumes. This means that non-phonetic characters may end up in the \"phonetic\" form, which we judge to be better than losing information through an inadequate mapping table.\n\n### Preprocesssors and postprocessors\n\nFor language-script pairs with more complicated orthographies, it is sometimes necessary to manipulate the orthographic form prior to mapping or to manipulate the phonetic form after mapping. This is done, in Epitran, with grammars of context-sensitive string rewrite rules. In truth, these rules would be more than adequate to solve the mapping problem as well but in practical terms, it is usually easier to let easy-to-understand and easy-to-maintain mapping files carry most of the weight of conversion and reserve the more powerful context sensitive grammar formalism for pre- and post-processing.\n\nThe preprocessor and postprocessor files have the same format. They consist of a sequence of lines, each consisting of one of four types:\n\n1. Symbol definitions\n2. Context-sensitive rewrite rules\n3. Comments\n4. Blank lines\n\n#### Symbol definitions\n\nLines like the following\n\n\n```\n::vowels:: = a|e|i|o|u\n```\n\n\ndefine symbols that can be reused in writing rules. Symbols must consist of a prefix of two colons, a sequence of one or more lowercase letters and underscores, and a suffix of two colons. The are separated from their definitions by the equals sign (optionally set off with white space). The definition consists of a substring from a regular expression.\n\nSymbols must be defined before they are referenced.\n\n#### Rewrite rules\n\nContext-sensitive rewrite rules in Epitran are written in a format familiar to phonologists but transparent to computer scientists. They can be schematized as\n\n\n```\na -> b / X _ Y\n```\n\n\nwhich can be rewitten as\n\n\n```\nXaY \u2192 XbY\n```\n\n\nThe arrow `->` can be read as \"is rewritten as\" and the slash `/` can be read as \"in the context\". The underscore indicates the position of the symbol(s) being rewritten. Another special symbol is the octothorp `#`, which indicates the beginning or end of a (word length) string (a word boundary). Consider the following rule:\n\n\n```\ne -> \u0259 / _ #\n```\n\n\nThis rule can be read as \"/e/ is rewritten as /\u0259/ in the context at the end of the word.\" A final special symbol is zero `0`, which represents the empty string. It is used in rules that insert or delete segments. Consider the following rule that deletes /\u0259/ between /k/ and /l/:\n\n\n```\n\u0259\u3000-> 0 / k _ l\n```\n\n\nAll rules must include the arrow operator, the slash operator, and the underscore. A rule that applies in a context-free fashion can be written in the following way:\n\n\n```\nch -> x / _\n```\n\n\nThe implementation of context-sensitive rules in Epitran pre- and post-processors uses regular expression replacement. Specifically, it employs the `regex` package, a drop-in replacement for `re`. Because of this, regular expression notation can be used in writing rules:\n\n\n```\nc -> s / _ [ie]\n```\n\n\nor\n\n\n```\nc -> s / _ (i|e)\n```\n\n\nFor a complete guide to `regex` regular expressions, see the documentation for [`re`](https://docs.python.org/2/library/re.html) and for [`regex`](https://pypi.python.org/pypi/regex), specifically.\n\nFragments of regular expressions can be assigned to symbols and reused throughout a file. For example, symbol for the disjunction of vowels in a language can be used in a rule that changes /u/ into /w/ before vowels:\n\n\n```\n::vowels:: = a|e|i|o|u\n...\nu -> w / _ (::vowels::)\n```\n\n\nThere is a special construct for handling cases of metathesis (where \"AB\" is replaced with \"BA\"). For example, the rule:\n\n\n```\n(?P<sw1>[\u0e40\u0e41\u0e42\u0e44\u0e43\u0e44])(?P<sw2>.) -> 0 / _\n```\n\n\nWill \"swap\" the positions of any character in \"\u0e40\u0e41\u0e42\u0e44\u0e43\u0e44\" and any following character. Left of the arrow, there should be two groups (surrounded by parentheses) with the names `sw1` and `sw2` (a name for a group is specified by `?P<name>` appearing immediately after the open parenthesis for a group). The substrings matched by the two groups, `sw1` and `sw2` will be \"swapped\" or metathesized. The item immediately right of the arrow is ignored, but the context is not.\n\n\nTo move IPA tones to the end of the word, first ensure that tones=True in the instantiated Epitran object and use the following rule:\n```\n(?P<sw1>[\u02e9\u02e8\u02e7\u02e6\u02e5]+)(?P<sw2>\\w+) -> 0 / _\\b\n```\n\n\nThe rules apply in order, so earlier rules may \"feed\" and \"bleed\" later rules. Therefore, their sequence is *very important* and can be leveraged in order to achieve valuable results.\n\n#### Comments and blank lines\n\nComments and blank lines (lines consisting only of white space) are allowed to make your code more readable. Any line in which the first non-whitespace character is a percent sign `%` is interpreted as comment. The rest of the line is ignored when the file is interpreted. Blank lines are also ignored.\n\n### A strategy for adding language support\n\nEpitran uses a mapping-and-repairs approach to G2P. It is expected that there is a mapping between graphemes and phonemes that can do most of the work of converting orthographic representations to phonological representations. In phonemically adequate orthogrphies, this mapping can do *all* of the work. This mapping should be completed first. For many languages, a basis for this mapping table already exists on [Wikipedia](http://www.wikipedia.org) and [Omniglot](http://www.omniglot.com) (though the Omniglot tables are typically not machine readable).\n\nOn the other hand, many writing systems deviate from the phonemically adequate idea. It is here that pre- and post-processors must be introduced. For example, in Swedish, the letter `<a>` receives a different pronunciation before two consonants (/\u0250/) than elsewhere (/\u0251\u02d0/). It makes sense to add a preprocessor rule that rewrites `<a>` as /\u0250/ before two consonants (and similar rules for the other vowels, since they are affected by the same condition). Preprocessor rules should generally be employed whenever the orthographic representation must be adjusted (by contextual changes, deletions, etc.) prior to the mapping step.\n\nOne common use for postprocessors is to eliminate characters that are needed by the preprocessors or maps, but which should not appear in the output. A classic example of this is the virama used in Indic scripts. In these scripts, in order to write a consonant *not followed* by a vowel, one uses the form of the consonant symbol with particular inherent vowel followed by a virama (which has various names in different Indic languages). An easy way of handling this is to allow the mapping to translate the consonant into an IPA consonant + an inherent vowel (which, for a given language, will always be the same), then use the postprocessor to delete the vowel + virama sequence (wherever it occurs).\n\nIn fact, any situation where a character that is introduced by the map needs to be subsequently deleted is a good use-case for postprocessors. Another example from Indian languages includes so-called schwa deletion. Some vowels implied by a direct mapping between the orthography and the phonology are not actually pronounced; these vowels can generally be predicted. In most languages, they occur in the context after a vowel+consonant sequence and before a consonant+vowel sequence. In other words, the rule looks like the following:\n\n\n```\n\u0259 -> 0 / (::vowel::)(::consonant::) _ (::consonant::)(::vowel::)\n```\n\n\nPerhaps the best way to learn how to structure language support for a new language is to consult the existing languages in Epitran. The French preprocessor `fra-Latn.txt` and the Thai postprocessor `tha-Thai.txt` illustrate many of the use-cases for these rules.\n\n# Citing Epitran\n\nIf you use Epitran in published work, or in other research, please use the following citation:\n\nDavid R. Mortensen, Siddharth Dalmia, and Patrick Littell. 2018. Epitran: Precision G2P for many languages. In *Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)*, Paris, France. European Language Resources Association (ELRA).\n\n```\n@InProceedings{Mortensen-et-al:2018,\n author = {Mortensen, David R. and Dalmia, Siddharth and Littell, Patrick},\n title = {Epitran: Precision {G2P} for Many Languages},\n booktitle = {Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)},\n year = {2018},\n month = {May},\n date = {7--12},\n location = {Miyazaki, Japan},\n editor = {Nicoletta Calzolari (Conference chair) and Khalid Choukri and Christopher Cieri and Thierry Declerck and Sara Goggi and Koiti Hasida and Hitoshi Isahara and Bente Maegaard and Joseph Mariani and H\\'el\\`ene Mazo and Asuncion Moreno and Jan Odijk and Stelios Piperidis and Takenobu Tokunaga},\n publisher = {European Language Resources Association (ELRA)},\n address = {Paris, France},\n isbn = {979-10-95546-00-9},\n language = {english}\n }\n ```\n",
"bugtrack_url": null,
"license": "MIT",
"summary": "Tools for transcribing languages into IPA.",
"version": "1.25",
"project_urls": {
"Download": "https://github.com/dmort27/epitran/archive/1.24.tar.gz",
"Homepage": "http://github.com/dmort27/epitran"
},
"split_keywords": [],
"urls": [
{
"comment_text": "",
"digests": {
"blake2b_256": "4015b9a7a3dedefc6c3883ffb7c7a5d2b1f67bdaa8f7136bff9876b58de46a7d",
"md5": "41205e8a5db8a68968b8984a355bd383",
"sha256": "178c09a04f0a5dbfce225b234de36f21e09145e9f4e6b06580cd495487eb7914"
},
"downloads": -1,
"filename": "epitran-1.25-py2.py3-none-any.whl",
"has_sig": false,
"md5_digest": "41205e8a5db8a68968b8984a355bd383",
"packagetype": "bdist_wheel",
"python_version": "py2.py3",
"requires_python": null,
"size": 173681,
"upload_time": "2024-03-08T09:27:36",
"upload_time_iso_8601": "2024-03-08T09:27:36.072839Z",
"url": "https://files.pythonhosted.org/packages/40/15/b9a7a3dedefc6c3883ffb7c7a5d2b1f67bdaa8f7136bff9876b58de46a7d/epitran-1.25-py2.py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "ed85e2a6b5ae4c734d30c25fd74dcbdc76af68b524de46c9659e8eb9bcd511e4",
"md5": "53c98dc9b06e2a46c53810396fcaca21",
"sha256": "740f248db9580891e07ea01222d8d57dd914f17c61148fda4c9fc102782bdbe9"
},
"downloads": -1,
"filename": "epitran-1.25.tar.gz",
"has_sig": false,
"md5_digest": "53c98dc9b06e2a46c53810396fcaca21",
"packagetype": "sdist",
"python_version": "source",
"requires_python": null,
"size": 126752,
"upload_time": "2024-03-08T09:27:38",
"upload_time_iso_8601": "2024-03-08T09:27:38.070341Z",
"url": "https://files.pythonhosted.org/packages/ed/85/e2a6b5ae4c734d30c25fd74dcbdc76af68b524de46c9659e8eb9bcd511e4/epitran-1.25.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2024-03-08 09:27:38",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "dmort27",
"github_project": "epitran",
"travis_ci": false,
"coveralls": false,
"github_actions": false,
"lcname": "epitran"
}
JSON
