Great Project, do you plan to add the VoiceClone model Qwen3-TTS-Base?

[id0o0bi]

Hi, This is a really great project, I can see many application of this
Do you have any plan to include more model from Qwen3-TTS?
like Base model for cloning voice, or maybe the 0.6B version.
or other models

[wavekat-eason]

Thanks for the suggestion! It’s a great idea—adding the Qwen3-TTS-Base model for voice cloning and the 0.6B version would definitely expand the project's capabilities.

I'm interested in looking into this; no problem at all. I'll start checking the implementation requirements and will look to add them soon. Stay tuned for updates!

[wavekat-eason]

Hi @id0o0bi,

Great news! We have just released v0.0.3 of wavekat-tts, which officially adds support for the Qwen3-TTS 0.6B Base model for voice cloning.

By default, the backend uses the INT4 precision model for efficiency, but we have also included support for FP32 precision if you require higher fidelity. The models will auto-download on the first run.

Voice Clone Example:

use wavekat_tts::AudioFrame;
use wavekat_tts::backends::qwen3_tts::{Qwen3TtsClone, CloneRequest};
// use wavekat_tts::backends::qwen3_tts::{ModelConfig, ModelPrecision};

fn main() {
    let ref_audio = AudioFrame::from_wav("ref.wav").unwrap();
    let ref_audio = ref_audio.resample(24000).unwrap(); // resample to 24 kHz (no-op if already 24 kHz)
    let tts = Qwen3TtsClone::new().unwrap(); // auto-downloads 0.6B Base INT4 model

    // For FP32 precision:
    // let config = ModelConfig::default().with_precision(ModelPrecision::Fp32);
    // let tts = Qwen3TtsClone::from_config(config).unwrap();

    let req = CloneRequest::new(
        "Text to say in the cloned voice",
        ref_audio.samples(),
        24000,
        "Transcript of the reference clip.",
    ).with_language("en");
    let audio = tts.synthesize_clone(&req).unwrap();
    audio.write_wav("clone_output.wav").unwrap();
    println!("Wrote clone_output.wav ({:.2}s)", audio.duration_secs());
}

[id0o0bi]

Hi @id0o0bi,

Great news! We have just released v0.0.3 of wavekat-tts, which officially adds support for the Qwen3-TTS 0.6B Base model for voice cloning.

Great Job 👍 , I tried the new model on my CPU. pretty good results for INT4
although sadly it's still pretty slow on CPU

[wavekat-eason]

Hi @id0o0bi,
Great news! We have just released v0.0.3 of wavekat-tts, which officially adds support for the Qwen3-TTS 0.6B Base model for voice cloning.

Great Job 👍 , I tried the new model on my CPU. pretty good results for INT4 although sadly it's still pretty slow on CPU

Glad to hear the INT4 results are working well for you!

Performance on the CPU is definitely a known limitation for these models at this stage. We might look into some optimizations if we have the time later on, but for now, we're just glad to hear the quality is meeting your expectations.

Thanks for sharing your feedback and for testing the new version!