Breaking

Open-source LuxTTS voice-cloning model runs at 150x realtime

July 4, 2026 at 05:10 EDT

Media Generation
Open Source

A lightweight open-source speech synthesis model called LuxTTS has been released. It supports zero-shot voice cloning that reproduces a speaker's voice from roughly three seconds of reference audio, and is said to generate 48kHz audio at more than 150 times realtime on a single GPU.

Open-Source Release · Apache-2.0

LuxTTS: voice cloning that fits in 1GB of VRAM

A lightweight text-to-speech model generating up to 150× realtime on GPU, running fully locally, with zero-shot voice cloning from just 3 seconds of reference audio.

150×

realtime generation on GPU — 150s of audio per second

48kHz

output sampling rate — clearer than the common 24kHz

<1GB

VRAM — runs on 4GB GPUs, CPU and Apple Silicon

reference audio for zero-shot voice cloning

Audio fidelity, drawn to scale

Sampling rate vs. many existing voice models.

24kHz

many models

48kHz

LuxTTS

2× the sampling rate — quality said to rival models 10× its size.

At a glance

Model size

~25–65 MB

Supported

CUDA · CPU · MPS (Mac)

Base

ZipVoice, distilled to 4 steps

License

Apache-2.0

PRAISED FOR

150× generation speed, faster-than-realtime on CPU
Fully local in under 1GB of VRAM
Clear 48kHz audio output
Practical cloning from ~3 seconds

NOTED LIMITATIONS

English-leaning cadence limits some languages
Occasional metallic artifacts (eased via return_smooth)
Pronunciation-vs-quality tradeoff needs t_shift tuning
Quality drops if the reference clip is too short

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.