OmniVoice-BG

A Bulgarian text-to-speech model, fine-tuned from k2-fsa/OmniVoice on Bulgarian audiobook narration. It produces natural, native-sounding Bulgarian speech, removing the foreign accent that stock OmniVoice has on Bulgarian.

This is the fine-tuned model weights only. It is one of three orthogonal pillars of the full OmniVoice-BG system — the other two (a lexical stress front-end and deterministic fixed-voice cloning) live in the OmniVoice-BG codebase and are applied at serving time, not baked into the weights. For correctly-stressed, consistent narration you want all three together.

Why it exists

Stock OmniVoice can speak Bulgarian (it is among its 600+ languages), but for narration three things were wrong out of the box:

1. A non-native "foreign" accent on the synthesized Bulgarian. → fixed by this fine-tuned model
2. Wrong lexical stress (e.g. БУтона instead of буТОна). → fixed by the stress front-end (serving-side)
3. A different voice every clip (voice design is non-deterministic). → fixed by deterministic fixed-voice cloning (serving-side)

Model details

• Architecture: OmniVoice — a diffusion/mask-prediction language model that predicts Higgs Audio V2 codec tokens (8 codebooks @ 24 kHz). Backbone is Qwen/Qwen3-0.6B.
• Base model: k2-fsa/OmniVoice (Apache-2.0)
• Fine-tuned on: raditotev/bg-audiobooks-tts (Bulgarian audiobook narration)
• Sample rate: 24 kHz mono
• Size: ~2.3 GB (model.safetensors, float32)

Training configuration

Setting	Value
init from	k2-fsa/OmniVoice
steps	3000
learning rate	1e-5, cosine, 3% warmup
batch	batch_tokens=1024, grad_accum=8
precision	bf16 mixed
attention	SDPA
weight decay	0.01
seed	42

Tuned to fine-tune on a single 16 GB GPU. See BG_FINETUNE_PLAN.md in the repo for the full pipeline (data prep → tokenize → train → A/B checkpoint selection).

Training metrics

Full TensorBoard logs are in tensorboard/ (rendered in the Metrics tab). Run summary over 3000 steps / 11 epochs:

• train/loss: 4.44 (step 25) → 3.99 (step 3000), min 3.88
• eval/loss: 4.07 (step 250) → 3.87 (step 2750, best) → 3.96 (step 3000)
• grad_norm: stable ~2.0–2.5 throughout; throughput: ~1.3 steps/sec

The released weights are the step-2750 checkpoint — the lowest validation loss, selected by eval loss and confirmed by A/B listening (eval_bg_ab.py).

eval step	250	500	1000	1250	1750	2250	2750	3000
eval/loss	4.07	4.13	4.07	3.92	3.94	3.95	3.87	3.96

Usage

Install the package and run inference with this model as a drop-in replacement for stock OmniVoice:

pip install omnivoice

omnivoice-infer \
    --model preslaff/omnivoice-bg \
    --text "Натиснете бутона, за да започнете." \
    --ref_audio narrator.wav \
    --output out.wav

Or from Python:

from omnivoice import OmniVoice

model = OmniVoice.from_pretrained("preslaff/omnivoice-bg", train=False)
wav = model.generate(
    text="Натиснете бутона, за да започнете.",
    ref_audio="narrator.wav",   # fixed reference voice for consistent cloning
)

For best results, mark Bulgarian lexical stress before passing text to the model (combining vowel + U+0301), and use a fixed reference clip with seeded greedy decoding. The OmniVoice-BG server (omnivoice/cli/serve.py) does both automatically.

License & attribution

Apache-2.0. This is a derivative of OmniVoice by k2-fsa, also Apache-2.0. The base model and architecture are © Xiaomi Corp. Fine-tuned for Bulgarian narration as part of the TrainGen project.