shadwl/voice-2

Per-language LoRA adapters for OmniVoice — Hmong, Bahnar, Khmer

Three low-rank adapters over a frozen multilingual OmniVoice base. Each adapter is ~114 MB; the 2.45 GB base is shared between all of them.

pip install -r requirements.txt

python infer_lora.py --adapter adapters/km --text "សួស្តី" --output out.wav \
    --ref_audio demo_voices/km_female.wav --ref_text "<the clip's transcript>"

Paths resolve relative to the repo, so nothing needs editing after cloning.

Adapters

dir	language	code	notes
adapters/km	Khmer	km	in the base model's training mix
adapters/bdq	Bahnar	bdq	in the base model's training mix
adapters/hmongv	Hmong	hmongv	new language, not seen by the base

hmongv is the Vietnam-based Latin orthography for Hmong (Vietnamese diacritics plus final-consonant tone letters, e.g. "Thâuv txos cheix ntux yiêz"). It is a different writing system from RPA White Hmong ("Txheeb xyuas cov kab lus"), which appears in the same source corpus under hmongz. The two are deliberately not merged — one tag for two orthographies would give the model contradictory spellings for identical sounds.

What is trained

component	params	trained
Qwen3 backbone (28 layers)	596M	frozen
LoRA r=16 on all attn + MLP projections	10.1M	yes
audio_embeddings (8200×1024)	8.4M	yes
audio_heads (1024×8200)	8.4M	yes
text embedding (151676×1024)	155M	frozen by default

26.9M / 639.5M trainable (4.2%). The audio embedding and head tables are trained in full because adapting to a new language is largely a matter of re-mapping which audio tokens follow which text, and a rank-16 update inside the backbone is a narrow channel for that. The text embedding is left frozen: it alone is nine times larger than everything else trainable combined, and Qwen3's byte-level BPE tokenizes unseen orthographies losslessly into known subwords. train_lora.py --train_text_embed turns it on.

Contents

train_lora.py            LoRA training (base frozen)
infer_lora.py            inference; load_lora_model() is the reusable entry point
hmong_to_lhotse.py       Hmong ASR corpus -> Lhotse Shar
parquet_to_lhotse.py     elego VC parquet -> Lhotse Shar
extract_tokens_parquet.py  Lhotse/parquet -> Higgs audio tokens
lhotse_dataset.py        shared Lhotse reader
parquet_dataset.py       shared parquet reader
registry.py              corpus registry
adapters/{km,bdq,hmongv}/  adapter weights + adapter_meta.json
base/                    frozen base checkpoint (inference weights)
demo_voices/             reference clips + transcripts

Reproducing

# 1. data -> Lhotse -> audio tokens
python hmong_to_lhotse.py --max_hours 150
python extract_tokens_parquet.py --lhotse_dir /path/lhotse/hmongv/train \
    --language_id hmongv --output_dir /path/tokens/hmongv/train --gpu_ids 0,1,2,3

# 2. one adapter per language, one GPU each
CUDA_VISIBLE_DEVICES=0 python train_lora.py --language hmongv \
    --output_dir exp/lora_hmongv --steps 8000

Notes and gotchas

• LoRA targets are a regex anchored on llm.layers.N., not bare module names. At inference OmniVoice attaches the Higgs audio tokenizer, whose encoder also has q_proj/k_proj/v_proj; bare names match those too and peft then injects untrained adapters into the audio codec.
• Ids must not contain dots. WebDataset splits a member name at the first dot to derive its key, so an id like hmongv_413024.0_00000002 is truncated and every label lookup fails — at training time, long after tokenization reports success.
• The Hmong training set excludes the ~44% of the source corpus that are synthetic augmentations (reverb / pitch-shift / time-mask). Those are fine as ASR inputs and ruinous as TTS targets — the model would learn to reproduce the reverb.
• Hmong speaker/gender metadata is largely absent, so its two reference clips are labelled voice1/voice2 rather than male/female.
• Use sdpa attention. flex_attention's Triton kernels fail on torch 2.7.1+cu118.

Base model

OmniVoice (Qwen3-0.6B + Higgs audio tokens) finetuned jointly on ~1,820 h across ten languages of Vietnam and Cambodia, 60k steps. Also published at shadwl/voice-demo.