MSpoofTTS

---

Hierarchical Decoding for Discrete Speech Synthesis with Multi-Resolution Spoof Detection

Paper | Demo | Checkpoints

Overview

Neural codec language models can synthesize high-quality speech, but inference in the discrete token space is still vulnerable to token-level artifacts and distribution drift. MSpoofTTS is a training-free inference framework that improves zero-shot speech synthesis with multi-resolution spoof guidance. It evaluates codec-token sequences at multiple temporal granularities to identify locally inconsistent or unnatural patterns, then integrates the resulting authenticity scores into hierarchical decoding to prune weak candidates and rerank hypotheses. The base NeuTTS language model is kept fixed; robustness is improved without retraining or changing its parameters.

News

• 2026.06: MSpoofTTS is accepted to INTERSPEECH 2026.
• 2026.04: arXiv v2 released: 2603.05373.
• 2026.03: Demo page with audio samples is available: MSpoofTTS demo.

Setup

git clone https://github.com/Danny-NUS/MSpoofTTS.git
cd MSpoofTTS

conda create -n mspooftts python=3.11 -y
conda activate mspooftts

# NeuTTS requires espeak-ng for phonemization.
brew install espeak-ng        # macOS
# sudo apt install espeak-ng  # Ubuntu/Debian

pip install -r requirements.txt

Optional backends:

pip install llama-cpp-python  # GGUF backbone support
pip install onnxruntime       # ONNX codec decoder support

Inference

The default example runs vanilla NeuTTS sampling:

python -m examples.basic_example \
  --input_text "MSpoofTTS helps neural text-to-speech systems generate more reliable speech by using spoof detection signals during decoding." \
  --ref_audio samples/jo.wav \
  --ref_text samples/jo.txt \
  --output_path generated/output.wav

Run MSpoofTTS-guided decoding by selecting rank_eas_hier. The discriminator checkpoints are downloaded automatically from Chanson-0803/MSpoofTTS:

python -m examples.basic_example \
  --input_text "MSpoofTTS is a training-free inference framework that improves discrete speech synthesis with multi-resolution spoof detection guided decoding." \
  --ref_audio samples/jo.wav \
  --ref_text samples/jo.txt \
  --sampling_scheme rank_eas_hier \
  --backbone_device cuda \
  --codec_device cuda \
  --output_path generated/output_mspooftts.wav

Python usage:

from neutts import NeuTTS
import soundfile as sf

tts = NeuTTS(
    backbone_repo="neuphonic/neutts-nano",
    backbone_device="cuda",
    codec_repo="neuphonic/neucodec",
    codec_device="cuda",
    use_hier=True,
    discriminator_repo="Chanson-0803/MSpoofTTS",
)

ref_text = open("samples/jo.txt", encoding="utf-8").read().strip()
ref_codes = tts.encode_reference("samples/jo.wav")

wav = tts.infer(
    "MSpoofTTS uses multi-resolution spoof scores during decoding.",
    ref_codes,
    ref_text,
    sampling_scheme="rank_eas_hier",
)
sf.write("generated/output_mspooftts.wav", wav, 24000)

Supported PyTorch decoding schemes:

What We Modified

This repository is based on NeuTTS. The upstream NeuTTS runtime is kept as the base system; MSpoofTTS additions are marked as (new) below and in code comments where they are mixed into existing files. The neutts/ directory is therefore modified upstream code, not a fully new package.

• (new) Discriminator/: multi-resolution token spoof detector modules used by guided decoding.
• (new) mspooftts/: Hugging Face checkpoint loading utilities for MSpoofTTS discriminator checkpoints.
• (new) rank_eas_hier in neutts/neutts.py: Entropy-Aware Sampling plus hierarchical discriminator pruning/reranking.
• (new) CLI flags in examples/basic_example.py: --sampling_scheme, --backbone_device, --codec_device, and --discriminator_repo.

New Code Mixed Into neutts/neutts.py

The original NeuTTS logic for phonemization, backbone loading, codec loading, reference encoding, decoding, streaming, and vanilla orig generation is kept as the base. The MSpoofTTS additions inside this file are:

• (new) imports for Discriminator and mspooftts.checkpoints.
• (new) nucleus_sampling and EASPenalty, used by EAS-based decoding.
• (new) optional constructor arguments: use_dis, use_hier, discriminator_repo, and discriminator_revision.
• (new) lazy discriminator loading when use_dis=True or use_hier=True.
• (new) speech-token helpers: _mask_to_speech_only and _lm_to_speech_id_or_none.
• (new) _eas_generate, _rank_sum_select, and _rank_eas_hier_generate.
• (new) sampling_scheme routing in infer / _infer_torch; orig remains the NeuTTS baseline path.

Repository Layout

TTSSpoofDetection/
├── assets/                 # (new) README figures and project icon
├── Discriminator/          # (new) MSpoofTTS discriminator modules used at inference
├── examples/               # Minimal command-line inference examples
├── generated/              # ignored local inference outputs
├── mspooftts/              # (new) MSpoofTTS checkpoint loading utilities
├── neutts/                 # (modified) NeuTTS runtime; new hooks marked in neutts.py
├── neuttsair/              # NeuTTS-Air compatibility wrapper
├── samples/                # Reference audio/text/code samples
├── tests/                  # Lightweight loading/inference smoke tests
├── README.md
├── LICENSE
├── LICENSE_APACHE
└── requirements.txt

Citation

If you use MSpoofTTS, please cite our paper. Since this repository builds on NeuTTS/NeuCodec components, please also cite NeuCodec when using the codec.

@article{zhao2026hierarchical,
  title={Hierarchical Decoding for Discrete Speech Synthesis with Multi-Resolution Spoof Detection},
  author={Zhao, Junchuan and Vu, Minh Duc and Wang, Ye},
  journal={arXiv preprint arXiv:2603.05373},
  year={2026}
}

@article{julian2025finite,
  title={Finite Scalar Quantization Enables Redundant and Transmission-Robust Neural Audio Compression at Low Bit-rates},
  author={Julian, Harry and Beeson, Rachel and Konathala, Lohith and Ulin, Johanna and Gao, Jiameng},
  journal={arXiv preprint arXiv:2509.09550},
  year={2025}
}

License

This repository contains two license scopes:

• NeuTTS-derived code and model usage: NeuTTS Open License v1.0 — see LICENSE.
• MSpoofTTS source code outside the original NeuTTS codebase, including Discriminator/, mspooftts/, README assets, and the marked (new) inference hooks in neutts/neutts.py: Apache-2.0 — see LICENSE_APACHE.

The MSpoofTTS discriminator checkpoints are hosted at Chanson-0803/MSpoofTTS. Please also follow the license and terms on that Hugging Face repository when using the checkpoints.