OmniVoice: the free Apache-2.0 ElevenLabs alternative that runs on your own GPU

#ai #news #nlp #opensource

Two AI launches in one day

While most people were arguing about Claude Sonnet pricing, two genuinely useful AI projects shipped — and they both deserve a closer look.

1. OmniVoice — the open-source ElevenLabs alternative

OmniVoice is from k2-fsa, the team behind Next-gen Kaldi (Xiaomi's speech research lab). It's Apache-2.0 licensed, supports 600+ languages, and runs at RTF 0.025 — that's 40 seconds of audio generated per 1 second of compute.

Three modes:

Voice Clone — clone any voice with just 3-10 seconds of reference audio
Voice Design — generate a voice from controls (gender, age, pitch, accent, dialect)
Auto Voice — let the model pick

Install:

pip install omnivoice

Use:

from omnivoice import OmniVoice

model = OmniVoice.from_pretrained("k2-fsa/OmniVoice")
audio = model.generate(
    text="Hello from a free voice model",
    ref_audio="reference.wav"
)

CLI tools that ship in the box: omnivoice-demo, omnivoice-infer, omnivoice-infer-batch.

If you want to try it without installing anything, the HuggingFace Space lets you generate samples in your browser.

Why this matters: ElevenLabs is the de facto API for AI voice, but it costs ~$11/month minimum and 600+ language coverage is genuinely difficult to match. OmniVoice flips both: free + broader coverage + local-first.

2. Google DreamBeans — the anti-infinite-scroll AI

DreamBeans launched June 3, 2026 from Google Labs. It connects to your Gmail, Calendar, Photos, YouTube, and Search history, then uses Nano Banana 2 (Google's latest image gen) to brew a finite daily collection of illustrated stories about your life.

The interesting part isn't the tech — it's the product decision. Google explicitly chose NOT to make it infinite scroll. You get a fixed number of stories per day, and that's it.

For now it's US-only on iOS/Android for Google AI Ultra subscribers. Everyone else can join the waitlist.