Downloads · 30 days
8
1% of all-time downloads
quannguyen204/VieNeu-TTS-0.3B-ngochuyen-elder-merged
VieNeu-TTS-0.3B-ngochuyen-elder-merged is a text-to-speech model from quannguyen204. Use it when you need text read aloud. It is set up for transformers. The card lists the license as cc-by-nc-4.0.
Drop-in merged model: base pnnbao-ump/VieNeu-TTS-0.3B + our LoRA adapter quannguyen204/VieNeu-TTS-0.3B-lora-ngochuyen-elder merged into a single bf16 checkpoint.
Downloads · 30 days
8
1% of all-time downloads
All-time downloads
945
Public
Parameters
244M
526 MB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors488 MB · 95%
From the Hugging Face model README
Drop-in merged model: base pnnbao-ump/VieNeu-TTS-0.3B + our LoRA adapter
quannguyen204/VieNeu-TTS-0.3B-lora-ngochuyen-elder
merged into a single bf16 checkpoint.
Use this repo when you want a single artifact (no PEFT dependency) for production / serving / quantization / mobile deployment. For development with hot-swappable adapters, use the LoRA-only repo above.
| Base model | pnnbao-ump/VieNeu-TTS-0.3B |
| LoRA r / alpha (merged in) | 16 / 32 |
| Max steps | 5000 (5.28 epochs) |
| Effective batch | 8 (per_device=4 × grad_accum=2, bf16) |
| Hardware | RTX 3090 Ti 24 GB (vast.ai) |
| Wall-clock | 1h 37m 30s |
| Final train loss | 1.099 |
| Final eval loss | 1.159 (gap=0.06, no overfit) |
| Source | Clips | Duration | Share |
|---|---|---|---|
| pnnbao-ump/ngochuyen_voice (HF, real) | 7238 | ~21.6h | 84.8% |
| Vbee elderly-dialogue synth | 1301 | ~1.2h | 15.2% |
| Train manifest total | 8539 | ~22.8h |
Pre-encoded dataset:
quannguyen204/tts-ngochuyen-encoded.
| Name | Role | Source |
|---|---|---|
NgocHuyenNews | formal news, real Ngọc Huyền | real HF clip |
NgocHuyenElder (default) | warm conversational, elderly companionship | Vbee synth |
NgocHuyenCare | gentle care instructions | Vbee synth |
from vieneu import Vieneu
tts = Vieneu(
mode="standard",
backbone_repo="quannguyen204/VieNeu-TTS-0.3B-ngochuyen-elder-merged",
backbone_device="cuda",
codec_device="cuda",
gguf_filename=None,
)
# voices.json shipped → no need to provide ref_codes manually
audio = tts.infer(
text="Dạ thưa bác, hôm nay là thứ Hai, ngày 26 tháng 5 năm 2026 ạ.",
voice={"name": "NgocHuyenElder"},
temperature=0.7,
top_k=40,
)
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "quannguyen204/VieNeu-TTS-0.3B-ngochuyen-elder-merged"
model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="cuda")
tokenizer = AutoTokenizer.from_pretrained(repo)
# Use NeuCodec for audio decoding — see neuphonic/neucodec.
On-robot Vietnamese voice assistant for elderly care/companionship: "dạ/vâng" persona, reminders, short conversational replies, time/date/weather, medication intake announcements.
sea_g2p PanicException workaround);
raw text conditioning may produce occasional pronunciation drift on rare
English / abbreviation tokens.CC BY-NC 4.0, inherits base model license. Non-commercial use only.