Crate qwen_tts 

Candle Qwen3-TTS

A Rust implementation of Qwen3-TTS text-to-speech model for the Candle ML framework.

This crate provides:

• High-level model API (model::Model)
• Speaker encoder (ECAPA-TDNN based)
• Audio tokenizer (12Hz)

Architecture Overview

Qwen3-TTS uses a hierarchical generation approach:

1. Text is processed through a talker model with multimodal RoPE
2. The first codebook (semantic) is predicted by the main talker
3. Remaining codebooks (acoustic) are predicted by a sub-talker (code predictor)
4. All 32 codes are decoded to audio via the tokenizer

Example

use qwen_tts::model::loader::{ModelLoader, LoaderConfig};
use candle_core::Device;

let loader = ModelLoader::from_local_dir("/path/to/model")?;
let model = loader.load_tts_model(&Device::Cpu, &LoaderConfig::default())?;
let result = model.generate_custom_voice_from_text(
    "Hello, world!",
    "vivian",
    "english",
    None,
    None,
)?;

Modules

audio

Audio processing components.

config

io

model

High-level TTS model wrapper.

nn

Model implementations for Qwen3-TTS.

synthesis

text

Text processing components.

Macros

increment_counter

Macro for incrementing call count. Compiles to no-op when timing feature is disabled.

record_elapsed

Macro for recording elapsed time. Compiles to no-op when timing feature is disabled.

timed

Macro for timing a code block. Compiles to no-op when timing feature is disabled.