flowty1/neutts-air-he

NeuTTS Air — Hebrew (LoRA)

A Hebrew LoRA finetune of neuphonic/neutts-air, adding natural Hebrew text-to-speech with instant voice cloning.

• Base model: neuphonic/neutts-air (Apache-2.0)
• Adapter type: LoRA (PEFT), rank 32, alpha 64, rsLoRA
• Language: Hebrew (he)
• Inference example / demo: flowtyone/neutts-air-he-lora on GitHub

This adapter must be applied on top of the base backbone; it does not contain the full model weights. For Hebrew text it is paired with the thewh1teagle/phonikud grapheme-to-phoneme (diacritization) model and neuphonic/neucodec.

Usage

import soundfile as sf
from peft import PeftModel
from transformers import AutoModelForCausalLM

# Minimal sketch — see the GitHub inference example for the full pipeline
# (phonemization with phonikud + audio codec with neucodec).
backbone = AutoModelForCausalLM.from_pretrained("neuphonic/neutts-air")
model = PeftModel.from_pretrained(backbone, "flowty1/neutts-air-he")

The full inference pipeline (reference encoding, phonemization, and waveform decoding) is provided in the GitHub repository, which loads this adapter directly by repo id:

from lib.neutts import NeuTTSAir

tts = NeuTTSAir(
    backbone_path="neuphonic/neutts-air",
    lora_path="flowty1/neutts-air-he",
    codec_path="neuphonic/neucodec",
    codec_encoder_path="facebook/w2v-bert-2.0",
    g2p_model_path="thewh1teagle/phonikud",
    device="auto",
)

Reference audio guidelines

For best voice-cloning quality, use a clean, mono reference clip of 3–15 seconds of natural, continuous speech, together with an accurate transcription.

Acknowledgements

• Neuphonic — NeuTTS Air and NeuCodec.
• Phonikud / DICTA — Hebrew diacritization.

License

This LoRA adapter is released under the Apache-2.0 license, matching the base model neuphonic/neutts-air.