Skip to main content

Phonemizers

KittenTTS converts normalized English text to IPA before inference.

Built-in CE phonemizer​

.builtin is the default. It uses the SDK's C++ engine and downloads en_rules and en_list dictionaries on first setup.

let config = KittenTTSConfig(phonemizer: .builtin)

Override dictionary download URLs by creating EPhonemizer directly:

let phonemizer = EPhonemizer(
    rulesURL: URL(string: "https://assets.example.com/en_rules")!,
    listURL: URL(string: "https://assets.example.com/en_list")!
)

let config = KittenTTSConfig(phonemizer: .custom(phonemizer))

eSpeak on macOS​

brew install espeak-ng
let config = KittenTTSConfig(phonemizer: .espeak)

Process execution is unavailable on iOS, so .espeak throws espeakNotInstalled there.

Custom implementation​

struct MyPhonemizer: KittenPhonemizerProtocol {
    func phonemize(_ text: String) -> String {
        customIPA(text)
    }
}

let config = KittenTTSConfig(phonemizer: .custom(MyPhonemizer()))

The returned IPA may only use symbols in the KittenTTS symbol table; unsupported scalars are discarded. Implement downloadIfNeeded(to:progressHandler:) when the custom engine needs its own assets.

Last updated on