---
license: cc-by-4.0
language:
- en
base_model: openbmb/VoxCPM-0.5B
base_model_relation: adapter
pipeline_tag: text-to-speech
tags:
- text-to-speech
- voxcpm
- lora
- accent
- voice
- vctk
- baby-mynah
---

# baby mynah — Welsh accent (LoRA for VoxCPM-0.5B)

A LoRA adapter that voices [VoxCPM-0.5B](https://huggingface.co/openbmb/VoxCPM-0.5B)
in the **Welsh** accent of a single VCTK speaker. One of six accent experts
behind the [**baby mynah** Space](https://huggingface.co/spaces/build-small-hackathon/baby-mynah), where
you can slide between any two of them in weight space.

## What it is

- **Base model:** [`openbmb/VoxCPM-0.5B`](https://huggingface.co/openbmb/VoxCPM-0.5B) (OpenBMB),
  a tokenizer-free TTS model. This repo holds **only the LoRA delta** (~28 MB) —
  you also need the base model.
- **Accent:** Welsh (Cardiff), learned from a single
  [VCTK](https://datashare.ed.ac.uk/handle/10283/3443) speaker (**p253**,
  age 22, female) — ~8 minutes of read speech.
- **Not PEFT.** VoxCPM uses a custom `LoRALinear`, so this loads through the
  baby-mynah loader, not `peft`. The weights file stores, per target Linear,
  `<prefix>.lora_A` (r, in) and `<prefix>.lora_B` (out, r); the layer applies
  `out += (alpha/r) * (x @ A^T @ B^T)`.

## Training

| | |
|---|---|
| Method | LoRA (rank 32, alpha 32, dropout 0) |
| Target modules | `q/k/v/o_proj` on both the LM and the DiT |
| Steps | 800 (effective batch 4 = batch 1 x grad-accum 4) |
| LR | 1e-4, 50 warmup, weight decay 0.01 |
| Audio | ~8 min, 16 kHz mono |

## Usage

These adapters are meant to be *blended*, which is what the Space does. To load
one on its own, use the project's `load_adapter` / `blend_into_model` helpers
(see [`scripts/blend.py`](https://huggingface.co/spaces/build-small-hackathon/baby-mynah/blob/main/scripts/blend.py)):

```python
from blend import load_adapter, blend_into_model, restore_weights, scaling_from_config
from voxcpm import VoxCPM

model = VoxCPM.from_pretrained(hf_model_id="openbmb/VoxCPM-0.5B", optimize=False, device="cuda")
adapter = load_adapter("path/to/this/adapter")        # {prefix: (A, B)}
scaling = scaling_from_config("path/to/this/adapter")  # alpha / r

# alpha=1.0 applies this expert fully; restore afterwards (surgery is in-place).
snap = blend_into_model(model.tts_model, adapter, adapter, alpha=1.0, scaling=scaling)
wav = model.generate(text="Hello! I am a very small bird.", normalize=True)
restore_weights(model.tts_model, snap)
```

## License & attribution

- LoRA weights released under **CC BY 4.0**, following the
  [VCTK corpus](https://datashare.ed.ac.uk/handle/10283/3443) (CSTR, University
  of Edinburgh) they were trained on.
- Base model **VoxCPM-0.5B** by **OpenBMB**, Apache-2.0.