huutamm1612/vieneu-tts-ngoc-huyen

VieNeu-TTS Ngọc Huyền

A full Vietnamese single-speaker TTS model adapted from pnnbao-ump/VieNeu-TTS.

The model was trained through three phases: LoRA adaptation, safe merge and partial fine-tuning. This repository contains the complete model, not a LoRA adapter.

Model information

• Base model: pnnbao-ump/VieNeu-TTS
• Codec: neuphonic/neucodec
• Dataset: pnnbao-ump/ngochuyen_voice
• Language: Vietnamese
• Code: huutamm1612/vieneu-ngoc-huyen-tts

Usage

Clone the training and inference repository:

git clone https://github.com/huutamm1612/vieneu-ngoc-huyen-tts.git
cd vieneu-ngoc-huyen-tts
python -m pip install -e ".[inference]"

Set HF_TOKEN through an environment variable or platform secret. The token must have access to the gated neuphonic/neucodec model.

from inference import InferenceConfig, TTSInference

config = InferenceConfig(
    model="huutamm1612/vieneu-tts-ngoc-huyen",
    devices="auto",
    num_gpus=1,
    max_runtime_batch_size=16,
    show_progress=True,
)

with TTSInference(config) as tts:
    result = tts.infer(
        input_path="input.txt",
        reference_audio="reference.wav",
        reference_text="Transcript phải khớp chính xác với reference audio.",
        output_path="output.wav",
    )

print(result.as_dict())

Recommended reference audio:

• 3–7 seconds
• Clean and continuous speech
• Minimal background noise
• Transcript matching the audio exactly

The pipeline automatically splits long Vietnamese text and combines all generated segments into one WAV file.

Limitations

The model works best with normalized Vietnamese narration similar to its training data. Foreign words, names, abbreviations, expressive speech and unusual symbols may be pronounced incorrectly.

MOS, WER and speaker-similarity evaluations have not yet been performed.

Responsible use

Use this model only for authorized voice synthesis. Do not use it for impersonation, fraud, deceptive content or presenting synthetic speech as an authentic recording.

License and credits

• Base model: pnnbao-ump/VieNeu-TTS, Apache-2.0
• Dataset: pnnbao-ump/ngochuyen_voice, CC BY-NC 4.0
• Codec: neuphonic/neucodec
• Code: huutamm1612/vieneu-ngoc-huyen-tts

This model is distributed under CC BY-NC 4.0 as conservative metadata due to the training dataset.