Most of the work doesn't need any programming knowledge, but, to get immediate feedback, by running and testing eSpeak, you should be able to build it.
You also have to understand the language main concepts, be aware of its features, and have to have patience and attention to detail. Wikipedia is a good source of basic phonetic information, e.g. about vowels and consonants.
In many cases it should be fairly easy to add a rough implementation of a new language, hopefully enough to be intelligible. After that it's a gradual process of improvement.
The language is identified using the BCP 47 language tag. The list of valid tags originate from various standards and have been combined into the IANA Language Subtag Registry.
These language tags are used to specify the language, such as:
fr (French) -- The ISO 639-1 2-letter language code for the language.
NOTE: BCP 47 uses ISO 639-1 codes for languages that are allocated 2-letter codes (e.g. using en instead of eng).
yue (Cantonese) -- The ISO 639-3 3-letter language codes for the language.
ta-Arab (Tamil written in the Arabic alphabet) -- The ISO 15924 4-letter script code.
NOTE: Where the script is the primary script for the language, the script tag should be omitted.
The languages are grouped by the closest language family the language belongs. These language families are defined in ISO 639-5. See also Wikipedia's List of language families for more details.
For example, the Celtic languages (Welsh, Irish Gaelic, Scottish Gaelic, etc.) are listed under the cel language family code.
If necessary, the language tags are also used to specify the accent or dialect of a language, such as:
es-419 (Spanish (Latin America)) -- The UN M.49 3-number region codes.
fr-CA (French (Canada)) -- Using the ISO 3166-2 2-letter region codes.
en-GB-scotland (English (Scotland)) -- This is using the BCP 47 variant tags.
en-GB-x-rp (English (Received Pronunciation)) -- This is using the bcp47-extensions language tags for accents that cannot be described using the available BCP 47 language tags.
NOTE: If the accent you are trying to describe cannot be specified using the above system, raise an issue in the bcp47-data project and a private use tag will be defined for that accent.
To add new language, you have to create or edit following files:
| path/file | action |
|---|---|
| Makefile.am | edit |
| espeak-ng-data/lang/roa/fr | create |
| phsource/phonemes | edit |
| phsource/ph_french | create (optional) |
| dictsource/fr_list | create |
| dictsource/fr_rules | create |
| dictsource/fr_extra | create (optional) |
where:
Makefile.am is build configuration file. It should be edited if you add new language text-to-phoneme translation rules or set of sounds (phonemes). Usually you have to add necessary entries for new language, but may not need if you add new dialect for existing language.
Search for configuration of existing languages (e.g. English) and add similar lines for your language in following sections. E.g. for French:
dictionaries: \ ... espeak-ng-data/fr_dict \ ... fr: espeak-ng-data/fr_dict espeak-ng-data/fr_dict: dictsource/fr_list dictsource/fr_rules dictsource/fr_extra dictsource/fr_emoji ...
Note, that you don‘t need to add fr_extra or fr_emoji reference in the last group, if your language doesn’t have this file.
Optionally, if you have phoneme definition file, following line also should be added:
phsource/phonemes.stamp: \ ... phsource/ph_french \ ...
E.g. espeak-ng-data/lang/roa/fr is the language file for French. This gives the language name and may set some options.
Each language needs a language file in espeak-ng-data/lang grouped by the language family. The filename of the default voice for a language should be the same as the language code (e.g. fr for French).
The simplest voice file would contain just 2 lines to give the language name (from the IANA Language Subtag Registry and language code, e.g.:
name French language fr
This language code specifies which phoneme table and dictionary to use (i.e. phonemetable fr and espeak-ng-data/fr_dict) to be used. If needed, these can be overridden by phonemes and dictionary attributes in the voice file. For example you may want to start the implementation of a new language by using the phoneme table of an existing language.
Details of the contents of language files are given in Voices.
Open file phsource/phonemes and add following lines
... phonemetable fr base ...
where: ... means other existing lines of the configuration file and base is usually is base1 or code of other (already defined) language:
base phoneme table contains only definitions of some “control” phonemes such as stress marks and pauses defined in top of the phonemes file.
base1 includes base and also has basic set of vowels and consonants defined under phonemetable base1 section in phonemes file.
base2 has additional phonemes (mostly diphones) included from ph_base2 file.
Other language e.g. en has all phonemes defined in its corresponding phoneme definition file e.g. phsourc/ph_english for English language.
... include ph_french ...
For example, if phsource/phonemes has following lines:
phonemetable fr base include ph_french
it defines a phoneme table fr which inherits phoneme definitions from base table and adds new definitions or overrides existing definitions of base phonemes from the phonemes/ph_french file.
If you decide to customize definitions of phonemes (vowel and consonant sounds) for the language, then you have to create e.g. phsource/ph_french phoneme definition file for French.
The phonemes file for the language defines those additional phonemes which are not inherited (generally the vowels and diphthongs, plus any additional consonants that are needed), or phonemes whose definitions differ from the inherited version (eg. the redefinition of a consonant).
Usually it will contain mostly vowels. Most consonants will be inherited from the common phoneme definitions in the master phoneme file: phsource/phonemes as described in Phonemes File section above.
Details of phonemes files are given in Phoneme Tables.
To build the phoneme files, run:
espeak-ng --compile-phonemes
or:
make phsource/phonemes.stamp
For many languages, the consonant phonemes which are already available in eSpeak, together with the available vowel files which can be used to define vowel phonemes, will be sufficient. At least for an initial implementation.
There are usually two dictionary files, e.g. for French:
dictsource/fr_list. This contains pronunciations for numbers, letter and symbol names, and words with exceptional pronunciations. It also gives attributes such as “unstressed” and “pause” to some common words. The fr_list file contains:
fr_rules, (e.g. foreign names).dictsource/fr_rules. This contains the spelling-to-phoneme translation rules.
Details of the contents of the dictionary files are given in Dictionary.
The fr_rules and fr_list files are compiled to produce the espeak-ng-data/fr_dict file, which eSpeak uses when it is speaking.
Once the language‘s phonemes have been defined, then pronunciation dictionary data can be produced in order to translate the language’s source text into phonemes. This consists of two source files: fr_rules (the spelling to phoneme rules) and fr_list (an exceptions list, and attributes of certain words). The corresponding compiled data file is espeak-ng-data/fr_dict which is produced from the fr_rules and fr_list sources by the command:
espeak-ng --compile=fr
or by:
make fr
All languages require automated tests to make sure accidental changes in program code don't change how the languages sound.
Tests are run with make check. It will print an error if a language is missing tests.
Most tests are bash scripts in tests/ and follow the syntax:
Depending on the test, expected result can be a phoneme string, an md5 hash or something else.
To add a test:
First add a test line with arbitrary expected result. Example line in tests/language-phonemes.test: test_phwav fi this_should_fail “ma na Na pa ba ta da”
run make check to see expected and actual results: testing fi 1c1 < this_should_fail
7541590a29e2392b827d9f1563e6579774fb15ea
The behaviour of the eSpeak program is controlled by various options such as:
Most of these options can be set in the voice file in espeak-ng-data/lang folder (see Voices). However some advanced options may be needed to add in source code in the function SelectTranslator() of the tr_languages.c file.
When espeak-ng is invoked with -X parameter, it shows more detailed trace of chosen language rules for pronunciation. This trace can show line numbers also, if language is compiled with --compile-debug option.
To do this, go to espeak-ng project root folder, then:
cd dictsource espeak-ng --compile-debug=en
When invoked in following way:
espeak-ng -ven -X "Test."
It will show:
Translate 'test' 1 5965: t [t] 1 2104: e [E] 1 5725: s [s] 1 5965: t [t] t'Est
Listen carefully to the eSpeak voice. Try to identify what sounds wrong and what needs to be improved.
Make the spelling-to-phoneme translation rules more accurate, including the position of stressed syllables within words. Some languages are easier than others. I expect most are easier than English.
Improve the sounds of the phonemes. It may be that a phoneme should sound different depending on adjacent sounds, or whether it's at the start or the end of a word, between vowels, in a stressed or unstressed syllable, etc. This may consist of making small adjustments to vowel and diphthong quality or length, or adjusting the strength of consonants. Phoneme definitions can include conditional statements which can be used to change the sound of a phoneme depending on its environment. Bigger changes may be recording new or replacement consonant sounds, or may even need program code to implement new types of sounds.
Some common words should be added to the dictionary (the fr_list file for the language) with an “unstressed” attribute $u or $u+ (e.g. in English, words such as “the”, “is”, “had”, “my”, “she”, “of”, “in”, “some”), or should be preceded by a short pause (such as “and”, “but”, “which”), or have other attributes, in order to make the speech flow better.
Improve the rhythm of the speech by adjusting the relative lengths of vowels in different contexts, e.g. stressed/unstressed syllable, or depending on the following phonemes. This is important for making the speech sound good for the language.
Make new intonation “tunes” for statements or questions (see Intonation).
For most of the eSpeak languages, I do not speak or understand the language, and I do not know how it should sound. I can only make improvements as a result of feedback from speakers of that language. If you want to help to improve a language, listen carefully and try to identify individual errors, either in the spelling-to-phoneme translation, the position of stressed syllables within words, or the sound of phonemes, or problems with rhythm and vowel lengths.