Lure Turbo

Lure Turbo 0.6B

The Lure voice, about 1.7x faster end to end.
_{Research preview · English · 0.6B parameters · 24 kHz}

Research preview. Weights, behaviour and this card can change between releases. Pin a revision if you build on it.

What Turbo changes

Lure, like every Qwen3-TTS 12Hz model, makes each 80 ms audio frame in two stages. A large model picks the first of the frame's 16 codes, then a small code predictor fills in the other 15 one after another. Once the large model is fast, those 15 sequential passes take most of the time per frame.

Lure Turbo replaces them with a head that predicts the 15 codes in 3 grouped passes (codes 1-3, 4-7 and 8-15). The size, voices, tags and directions are the same as Lure's.

Per frame, 24 GB GPU, 8 streams	Lure	Lure Turbo
Code predictor passes	15	3
Code predictor time	21.2 ms	6.2 ms

Speed

Measured with the lure-tts engine, cuda backend, bf16. RTF is generation time divided by audio length, so lower is faster.

GPU	Model	Streams	RTF
H100	Lure	1	0.23
H100	Lure Turbo	1	0.14
H100	Lure	8	0.25
H100	Lure Turbo	8	0.15

24 GB GPU	1 stream	8 streams, total
Lure	2.3x real time	17x real time
Lure Turbo	4.0x real time	30x real time

Quality

On every public benchmark we ran, Turbo scores the same as Lure (one take per line, official scripts):

	Seed-TTS WER ↓	Seed-TTS SIM ↑	LibriSpeech-PC WER ↓	LibriSpeech-PC SIM ↑	UTMOS ↑	VDB SIM ↑	CV3 emotion % ↑
Lure	1.51	0.653	1.72	0.638	4.43	0.81	73.3
Lure Turbo	1.59	0.654	1.84	0.641	4.38	0.81	73.3

The full comparison with other models is on the Lure card.

Quick start

The grouped head needs the lure-tts engine, which loads it automatically when turbo.json is in the model folder (mlx, cuda and C++ backends). Other Qwen3-TTS runtimes do not know the head.

git clone https://github.com/smirki/lure-tts && cd lure-tts
hf download Tesslate/Lure-TTS-Turbo --local-dir models/lure-turbo
pip install ./backends/mlx ".[server]"        # Mac; see the repo for NVIDIA and the C++ build

from lure_tts import Lure

tts = Lure.load("models/lure-turbo")
voice = tts.make_voice("reference.wav", "Transcript of the reference clip.", name="me")
for pcm in tts.stream("(gasp) No way. (laugh) You actually did it!", voice, instruct="Bright, giddy disbelief."):
    ...    # float32 chunks at 24 kHz

Tags, directions and recommended settings are the same as Lure's; see the Lure card.

Limitations

Same as Lure: English only, sound tags are best effort, and very long single calls should be split into sentences. Do not use it to imitate a real person without their consent.

Training data includes YODAS v3.