Warble is a small gift from ScrappyLabs. It's five named voices living in one tiny model: just say which one you want and it talks — no reference clip, no cloud, no account, no meter. It runs on a laptop in real time, and it runs on an Apple-Silicon phone-class chip too.

Credit where it's due

Warble is built entirely on Audio8/Audio8-TTS-Preview-0.6b by the Audio8 team (Apache-2.0). The clever part — a capable 0.6B text-to-speech model — is theirs. All we did was train a few voices we liked onto it and package it so it's easy to run. If you like Warble, go star their work first; none of this exists without them.

The voices

Pick a voice by putting its token at the front of your text. No reference audio needed.

token	voice	character
`<\|speaker:1\|>`	Silas	deep, gravelled, movie-trailer weight
`<\|speaker:2\|>`	Narrator	even, warm, broadcast baritone
`<\|speaker:3\|>`	Clara	clear, professional, friendly
`<\|speaker:4\|>`	Pip	bright, light, quick
`<\|speaker:5\|>`	Nova	confident morning-radio energy

Samples for each are in samples/.

Use it

python

from transformers import AutoProcessor, AutoModel
import torch, soundfile as sf

model_id = "scrappylabsai/warble"
proc  = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModel.from_pretrained(model_id, trust_remote_code=True, dtype=torch.bfloat16).to("cuda").eval()

text = "<|speaker:2|>Hi there. This is Warble, running right here on your own machine."
inputs = proc(text=[text], return_tensors="pt")
inputs = {k: (v.to("cuda") if hasattr(v, "to") else v) for k, v in inputs.items()}
wavs, lens, _ = model.generate_audio(**inputs, max_new_tokens=1024)
sf.write("out.wav", wavs[0][:int(lens[0])].float().cpu().numpy(), 44100)

On a Mac / phone: an MLX 4-bit build is included — it runs on Apple Silicon, including phone-class chips, in about the memory a couple of browser tabs use.

Honest notes

Our voices are consent-first. Silas, Narrator, Clara, Pip, and Nova were designed by us from scratch — not cloned from any real person, not scraped from anyone. We can't make that claim about the base model's own training data (that's Audio8's and undocumented), so we won't — but the five voices we're handing you are entirely our own.
The MLX 4-bit build trades a little voice-likeness for size; Silas and Clara soften the most there. Use the full-precision model if you want them exactly right.
It's a preview-generation base model, so expect the occasional rough edge.

License

Apache-2.0, matching the base. Use it for anything, including commercial work. No strings.

Made by ScrappyLabs. We make things work, and better. This one's on the house. 🛠️