Text-to-Speech Overview¶

Text-to-speech (TTS) synthesis converts text to audio using neural models.

Requirements¶

TTS requires the mod-speech feature:

export LIBTORCH=/opt/homebrew  # macOS
# or
export LIBTORCH=/opt/libtorch   # Linux

cargo build --features mod-speech
divvun-runtime sync

Pipeline Flow¶

Text → Tokenize → Normalize → Phonology → Sentences → TTS → Audio

Tokenize - Split text into morphological units (hfst.tokenize)
Normalize - Convert numbers, abbreviations to words (speech.normalize)
Phonology - Add phonological forms (speech.phon)
Sentences - Extract sentence strings (cg3.sentences)
TTS - Synthesize speech (speech.tts)

Speech Modules¶

speech.normalize¶

Normalize text for speech:

let x = speech.normalize(input, {
    normalizers: { "Sem/Plc": "place-norm.hfst" },
    generator: "generator.hfst",
    analyzer: "analyzer.hfst"
});

Converts: - Numbers to words (123 → "one hundred twenty-three") - Abbreviations to full forms (Dr. → "Doctor") - Special tags to appropriate forms

speech.phon¶

Add phonological representations:

let x = speech.phon(input, {
    model: "phon.hfst",
    tag_models: { "Prop": "phon-prop.hfst" }
});

Generates pronunciation forms for synthesis.

speech.tts¶

Synthesize audio:

let audio = speech.tts(sentences, {
    voice_model: "voice.onnx",
    univnet_model: "vocoder.onnx",
    speaker: 0,
    language: 0,
    alphabet: "sme"  // "sme", "smj", "sma", "smi"
});

Returns WAV audio bytes.

Sentence Extraction¶

Extract sentences with phonological forms:

let sentences = cg3.sentences(input, {
    mode: "phonological"
});

Returns array of sentence strings ready for TTS.

Supported Languages¶

Current alphabet options: - sme - Northern Sami - smj - Lule Sami - sma - Southern Sami - smi - Generic Sami

Audio Output¶

TTS returns bytes (WAV format). Save to file:

divvun-runtime run -o output.wav bundle.drb "text to speak"

Or process further in TypeScript.