baby mynah is a small TTS demo built for the Hugging Face Build Small Hackathon: on top of VoxCPM-0.5B, it trains one LoRA expert per English accent and morphs smoothly between two accents via weight-space interpolation. In the author's words, it is an animal with an identity crisis.
A bird that can't pick an accent
Mynahs are nature's great mimics — they collect voices the way other birds collect twigs. This chick has heard nine English accents (English, Indian, American, Scottish, South African, New Zealand, Welsh, Irish, Northern Irish — all real human voices from VCTK) and can't pick one.
Give her a line, choose a voice for each ear, and slide her between them.
How she works
- Base model: VoxCPM-0.5B by OpenBMB — a tokenizer-free TTS model. ~0.73B parameters loaded in total (652M base + 75M audio VAE). Built small on purpose.
- One LoRA expert per accent, fine-tuned on a single VCTK speaker (~8 min of read speech each).
- The morph slider is weight-space surgery, not crossfading: for each target layer the two experts' effective deltas are blended as
W = W_base + s·[α·(B_a@A_a) + (1−α)·(B_b@A_b)]and patched into the base weights for that render, then restored. α=1 is all of one accent, α=0 all of the other, and everything in between is a voice that has never existed. - Renders are seeded from the text + accent pair (α excluded), so a slider sweep changes only the morph — not the random performance.
She speaks English only — accents, not languages. She ranks nothing and corrects no one; every voice is a whole identity and she simply tries them all on.
Credits
- VoxCPM-0.5B — OpenBMB (model · code, Apache-2.0)
- VCTK corpus — CSTR, University of Edinburgh (CC BY 4.0)
- Built for the Hugging Face Build Small Hackathon
Links
- Hugging Face Space: build-small-hackathon/baby-mynah
- Demo video: YouTube
- Launch post: on X
- Author: @tanenyao