Introduction

Lure is a new open text-to-speech model from Tesslate, and it landed on Hugging Face only hours ago in two versions: Lure 0.6B and Lure Turbo. Both are English-only voice cloners with 0.6 billion parameters under the Apache 2.0 license. They come with their own runtime, lure-tts, which is also Apache 2.0.

Here's the pitch in one line from the model card: it "laughs, sighs, breaks and whispers where you tell it to." You give it a few seconds of someone's voice plus the words they said, then write your script with sound cues inline, like (sigh) or (sharp inhale). Lure performs those sounds at that spot instead of reading them out loud. You can also add a short acting note, and it changes how the line is delivered.

Tesslate labels both models a research preview, so the weights and the card may change between releases. An earlier preview checkpoint had been floating around without a license this week. This is the first official release with clear terms.

What It Does

  • Voice cloning: zero-shot from a 3 to 12 second reference clip plus its exact transcript. Voices are encoded once and cached by name.
  • Sound tags: 36 inline tags in round brackets, grouped into laughter, breath, crying, throat and mouth sounds, spoken sounds like (hmm) and (ugh), and (pause). Square brackets like [laugh] don't work.
  • Directions: an optional free-text note to the "actor," such as "Hushed and shaking, close to tears." This needs the lure-tts runtime.
  • Output: 24 kHz mono audio that streams in chunks of about 0.32 seconds while the rest is still being generated.
  • Format: it uses the Qwen3-TTS 12Hz codec format (12.5 Hz, 16 codebooks), so Alibaba's qwen-tts package can load it directly too, just without Turbo or directions.

The card is honest about the rough edges. Laughs, sighs, gasps and sharp inhales land most reliably, while rarer tags like (ow) or (sob) get skipped more often. Lure is also weaker than big instruction-tuned models at changing who is speaking, like accent, age or gender, on request.

Lure vs. Lure Turbo

Every Qwen3-TTS-style model builds each 80 ms audio frame in two stages. A large model picks the first of 16 codes, then a small predictor fills in the other 15 one at a time. Turbo swaps that predictor for a head that fills the 15 codes in 3 grouped passes instead of 15.

Tesslate's speed numbers, measured with lure-tts in bf16:

  • H100, one stream: Lure runs at an RTF (real-time factor, generation time divided by audio length, lower is faster) of 0.23. Turbo runs at 0.14.
  • 24 GB GPU, one stream: Lure makes audio 2.3 times faster than real time. Turbo hits 4 times.
  • 24 GB GPU, eight streams at once: Lure totals 17 times real time and Turbo totals 30 times.
  • Mac: the card says Lure runs faster than real time on a MacBook.

Same size, same voices, same tags. Turbo is simply the one to pick if speed matters.

The Benchmarks

Tesslate ran Lure through each benchmark's official scripts, one take per line with default sampling. The rows for other models are the numbers their own developers published, so treat this as a self-reported comparison.

  • Naturalness (LibriSpeech-PC UTMOS, higher is better): Lure scores 4.43, the top value in Tesslate's table. That's ahead of Qwen3-TTS at 4.41, Fish Drama 3 at 4.31, and CosyVoice3 and OmniVoice at 4.28. Real recordings score 4.10 on this metric.
  • Word accuracy (Seed-TTS English WER, lower is better): Lure gets 1.51% and Turbo 1.59%. Fish S2 Pro (5B) still wins at 0.99% and Qwen3-TTS 0.6B posts 1.32%, but Lure edges out Fish Drama 3 (35B MoE) at 1.54% and VoxCPM2 (2B) at 1.84%.
  • Voice similarity (Seed-TTS SIM, higher is better): Lure scores 0.653. That's behind FireRedTTS3 at 0.772 and VoxCPM2 at 0.753, so cloning fidelity is not where it leads.
  • Emotion (CV3 emotion accuracy): 73.3%, close to Ming Tiny at 74.0% and behind CosyVoice3 1.5B with emotion tuning at 77.7%.

The short version: it sounds very natural and reads text accurately for its size, while bigger models still clone a specific voice more closely.

Why It Matters for Audio Creators

Non-speech sounds are what make a read feel acted rather than narrated. A sigh before a confession, a caught breath before a reveal, a nervous laugh mid-sentence. Until recently, that kind of control mostly lived in paid tools like ElevenLabs' v3 audio tags or in much heavier open models.

Lure puts it in a 0.6B package with a permissive license. That's small enough to run locally for audio drama drafts, game dialogue temp tracks, character voice tests or podcast inserts, with no per-character bill. Apache 2.0 means you can build it into commercial projects too. The built-in server speaks OpenAI's POST /v1/audio/speech API, so tools already wired for that can point at your own machine.

Do keep the card's own warning in mind: only clone a real person's voice with their consent. The training data includes the YODAS v3 speech corpus.

How to Run It Today

On a Mac with Apple silicon:

git clone https://github.com/smirki/lure-tts && cd lure-tts
hf download Tesslate/Lure-TTS-0.6B --local-dir models/lure
pip install ./backends/mlx ".[server]"
export LURE_MODEL=models/lure
lure say "Wait, you're staying? (sharp inhale) Since when?" --ref-wav ref.wav --ref-text "Transcript of ref.wav." --save-voice me -o out.wav

  • NVIDIA GPUs: the cuda backend batches up to 8 requests at once. Run lure serve --port 8880 for the OpenAI-compatible server. The repo has the exact install steps.
  • CPU and other GPUs: a C++ build covers CPU, Metal, CUDA and Vulkan on macOS, Linux and Windows.
  • Turbo: download Tesslate/Lure-TTS-Turbo instead and point the same commands at it.
  • Settings that matter: always set the language to English, keep one or two tags per sentence, and use a clean single-speaker reference clip with no music. If a take misses a tag or runs long, regenerate with a new seed.

Conclusion

Lure is a fun, practical drop. It's a tiny open voice cloner that takes stage directions, sounds more natural than models many times its size on Tesslate's tests, and ships with a real runtime for Mac, NVIDIA and CPU on day one. It's English-only and still a preview, and bigger models clone a specific voice more faithfully. But if you've wanted to script a sigh, a gasp or a laugh into a local TTS pipeline, this is the one to try first. Watch for non-preview weights and independent arena results.

—Anabel ♡