You are viewing the documentation for a prerelease version. View Latest

Usage

This page walks through `dengjen-piper-rs’s public API: loading a model and synthesizing audio from text. For the full model lifecycle (load, unload, swap voices), see Model lifecycle. For runnable programs, see Examples.

Loading a model

Piper::new takes the paths to a Piper .onnx model and its .onnx.json config, and returns a PiperResult<Piper>:

use dengjen_piper_rs::Piper;
use std::path::Path;

let mut piper = Piper::new(
    Path::new("path/to/voice.onnx"),
    Path::new("path/to/voice.onnx.json"),
)?;

Loading fails with PiperError::FailedToLoadResource if the config can’t be opened or parsed, or if the ONNX Runtime session can’t be created from the model file.

If you’ve already built an ort::Session yourself — e.g. with custom execution providers — use Piper::from_session instead, passing the parsed ModelConfig:

use dengjen_piper_rs::{ModelConfig, Piper};

let piper = Piper::from_session(session, config);

from_session doesn’t return a Result: it can’t fail the way new can, since the session is already built.

Synthesizing speech

Piper::create turns text (or phonemes) into raw audio samples:

let (samples, sample_rate) = piper.create(
    "Hello! I'm playing audio from memory directly with piper-rs.",
    false,        // is_phonemes
    None,         // speaker_id
    None,         // length_scale
    None,         // noise_scale
    None,         // noise_w
)?;

It returns (Vec<f32>, u32): mono samples in [-1.0, 1.0], and the sample rate declared in the model’s config (audio.sample_rate).

Parameters:

text — the input string. Interpreted as plain text or as raw phonemes, depending on is_phonemes.
is_phonemes — when true, text is fed to the model’s phoneme encoder directly, skipping the phonemizer backend entirely. Use this if you’ve already produced phonemes some other way.
speaker_id — selects a speaker on multi-speaker models. Ignored (single implicit speaker) on single-speaker models. See Multi-speaker voices below for how to find valid IDs.
length_scale, noise_scale, noise_w — override the model’s inference defaults from its .onnx.json config. None keeps the config’s default for that parameter.

create fails with PiperError::PhonemizationError if the phonemizer backend can’t process the text, or PiperError::InferenceError if the ONNX Runtime session fails.

Playing audio

examples/usage.rs plays the returned samples directly with rodio:

let device = rodio::DeviceSinkBuilder::open_default_sink().unwrap();
let player = rodio::Player::connect_new(device.mixer());
let channels = NonZero::new(1).unwrap();
let sample_rate = NonZero::new(sample_rate).unwrap();
player.append(SamplesBuffer::new(channels, sample_rate, samples));
player.sleep_until_end();

Saving audio

examples/wav.rs converts the f32 samples to 16-bit PCM and writes a WAV file — see Examples for the full listing.

Multi-speaker voices

Piper::voices reports whether the loaded model is multi-speaker:

match piper.voices() {
    None => println!("Single-speaker model."),
    Some(voices) => {
        for (name, id) in voices {
            println!("ID: {id}  Name: {name}");
        }
    }
}

It returns None for single-speaker models, or Some(&HashMap<String, i64>) mapping speaker name to speaker ID for multi-speaker ones. Pass one of those IDs as create’s `speaker_id argument to select a speaker.

Choosing a phonemizer backend

dengjen-piper-rs needs exactly one of two mutually exclusive Cargo features enabled to turn text into phonemes:

espeak-rs (default) — pure-Rust eSpeak NG bindings via dengjen-espeak-rs-adapter.
espeak-ng — the espeak-ng crate. When this feature is active, set the PIPER_ESPEAKNG_DATA_DIRECTORY env var to the directory containing espeak-ng-data if it isn’t found automatically.

Enabling both, or neither, fails to compile.

Want to help? Learn how to contribute to the ZirekHQ docs ›