nguyen-brat/vieneu-tts-vi

VieNeu-TTS Vietnamese LoRA Adapter

LoRA adapter (r=32, α=64) for pnnbao-ump/VieNeu-TTS, fine-tuned on ~80k segments of Vietnamese audiobook speech.

For most users, the merged model is easier to use: 👉 nguyen-brat/VieNeu-TTS-Vietnamese-Finetuned

🎙️ Live demo: nguyen-brat/vieneu-tts-demo

---

Usage

With VieNeu SDK (recommended)

from vieneu import Vieneu

# Load base model and apply LoRA adapter
tts = Vieneu(backbone_repo="pnnbao-ump/VieNeu-TTS")
tts.load_lora_adapter("nguyen-brat/vieneu-tts-vi")  # HF Hub repo ID

# Generate speech
audio = tts.infer("Xin chào, đây là giọng nói tiếng Việt.")

import soundfile as sf
sf.write("output.wav", audio, samplerate=24000)

Voice Cloning

from vieneu import Vieneu

tts = Vieneu(backbone_repo="pnnbao-ump/VieNeu-TTS")
tts.load_lora_adapter("nguyen-brat/vieneu-tts-vi")

audio = tts.infer(
    text="Đây là giọng được nhân bản từ đoạn âm thanh tham chiếu.",
    ref_audio="reference.wav",     # 3–10 second WAV of target speaker
    ref_text="Transcript here.",   # optional but recommended
)

With PEFT directly (backbone only)

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("pnnbao-ump/VieNeu-TTS")
model = PeftModel.from_pretrained(base, "nguyen-brat/vieneu-tts-vi")
tokenizer = AutoTokenizer.from_pretrained("nguyen-brat/vieneu-tts-vi")

# Note: This only loads the LM backbone. For full TTS inference
# (text → audio) you need the VieNeu SDK for phonemization and codec decoding.

---

Adapter Details

Parameter	Value
PEFT type	LoRA
Rank (r)	32
Alpha	64
Dropout	0.05
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Task type	CAUSAL_LM

Training

See nguyen-brat/VieNeu-TTS-Vietnamese-Finetuned for full training details.

License

Apache 2.0