Languages and data

espeak-ng supports over a hundred languages and accents. Each one is a voice file (tiny) plus a compiled dictionary (anything from 2 KB to 8 MB), and both live in the espeak-ng-data directory this package ships.

Voices

listVoices() returns every voice espeak-ng knows about, whether or not its dictionary is loaded yet:

for (const voice of espeak.listVoices()) {
  console.log(voice.name, voice.languages.map((l) => l.name).join(", "));
}
// English (America)  en-us
// German             de
// ...

Select one by name or language tag (setVoice("de"), setVoice("en-gb")), or by properties:

espeak.setVoice({ language: "en", gender: "female" });
espeak.setVoice({ name: "English (Received Pronunciation)" });

A string is tried as a voice name first and then as a language tag, which has to be one a voice lists (see listVoices), though case doesn't matter. When several voices list it, the one that gives it the best priority wins. Anything else throws EspeakError, so a typo fails instead of landing on some other voice. The language property is looser: espeak-ng picks the closest voice it has, so { language: "en-us-x-anything" } still finds en-us.

Variants

espeak-ng has voice variants (different pitches, speeds, and the famous "whisper") in espeak-ng-data/voices/!v. Select one with the + syntax espeak-ng uses on the command line:

espeak.setVoice("en-us+f3"); // female variant 3
espeak.setVoice("de+whisper");

Dictionaries

Before a language can be spoken or transcribed, its dictionary has to be in the module's virtual file system.

• Node.js reads dictionaries from disk the first time a voice needs one. Nothing to do.
• Browsers fetch them. English is fetched by createEspeak; add others with the languages option, loadLanguages() or loadVoice().

availableLanguages() lists what the data source can provide, loadedLanguages() what is ready to use. Language codes are espeak-ng's dictionary names: en, de, cmn, pt, es, ru and so on. They are the first part of the voice's language tag in most cases, though not all: en-gb and en-us share en, and es-419 uses es.

if (!espeak.loadedLanguages().includes("fr")) {
  await espeak.loadLanguages(["fr"]);
}

The data directory

espeak-ng-data/
  manifest.json   lists everything below, with sizes
  core.bin.gz     phoneme tables, intonations, voice and language definitions
  af_dict.gz      one gzipped dictionary per language
  am_dict.gz
  ...

core.bin.gz is every file except the dictionaries, concatenated in the order manifest.json lists them, so a browser makes one request for the 250 or so small files instead of 250. All files are gzipped at build time and decompressed by the library with DecompressionStream, which every supported runtime has.

File	Compressed	Decoded
core.bin.gz	349 KB	725 KB
en_dict.gz	108 KB	290 KB
ru_dict.gz	4.8 MB	8.1 MB
all 114 languages	9.1 MB	19 MB

The manifest's encoding field says how the files are stored; a repacked directory with "identity" (uncompressed) files is also accepted.

Bringing your own data

The loaders only need a DataSource: the manifest plus a function that returns a file's decoded bytes. Anything that can produce those works, such as an IndexedDB cache or files bundled into an app. See the low-level API.