neutts-nano-q8-gguf

NeuTTS Nano Q8 GGUF (English)

Q8_0 GGUF backbone for NeuTTS Nano (English), hosted under aoiandroid for local / on-device voice cloning in [on-device iOS client](local iOS project (private)) and compatible tooling.

[!NOTE] This checkpoint is English only. Use a French / German / Spanish sibling repo for those languages, or see the NeuTTS Nano Multilingual Collection.

Upstream

Weights in this repo follow the same Q8_0 layout and filename convention as the upstream neuphonic/neutts-nano-q8-gguf release (~241 MB backbone). This is a distribution mirror, not a new architecture.

Model description

NeuTTS Nano is a compact speech language model + NeuCodec stack for instant voice cloning: a few seconds of reference audio plus its transcript condition a new utterance. The GGUF file here is only the LM backbone; full synthesis still needs the NeuCodec decoder (see upstream neutts).

• Quantization: Q8_0 (GGUF)
• Context length: 2048 tokens (Nano family; see upstream docs)
• Output audio: 24 kHz via NeuCodec decode

Intended use

• On-device or local-first TTS with voice cloning (assistants, dubbing experiments, privacy-sensitive workflows)
• macOS integrations that load this repo id as the GGUF backbone (e.g. on-device iOS client NeuTTSNanoGGUFService, backbone code en)

Not intended for: unattended public APIs without safety review, impersonation, or use without rights to cloned voices.

Languages

• Primary: English (en)
• Phonemizer / espeak-ng locale: en when using the official neutts CLI

Use same-language reference audio and transcript as the backbone language.

Files in this repository

.
├── .gitattributes
├── README.md
└── neutts-nano-Q8_0.gguf    # Q8_0 LM backbone (~241 MB)

How to run

Recommended: neutts (Python)

1. Install espeak-ng (required for phonemization).
2. Install NeuTTS with GGUF support:

pip install "neutts[llama]" soundfile

1. Synthesize with this repo as the backbone (downloads codec from Hub on first run):

from neutts import NeuTTS
import soundfile as sf

tts = NeuTTS(
    backbone_repo="aoiandroid/neutts-nano-q8-gguf",
    backbone_device="cpu",
    codec_repo="neuphonic/neucodec",
    codec_device="cpu",
)

ref_text = open("reference.txt", "r", encoding="utf-8").read().strip()
ref_codes = tts.encode_reference("reference.wav")

wav = tts.infer("Your english text here.", ref_codes, ref_text)
sf.write("out.wav", wav, 24000)

CLI example (from the neutts repo):

python -m examples.basic_example \
  --input_text "Your sentence." \
  --ref_audio reference.wav \
  --ref_text reference.txt \
  --backbone aoiandroid/neutts-nano-q8-gguf

Local GGUF path

After downloading neutts-nano-Q8_0.gguf, you can point NeuTTS at the file path instead of the Hub id if your integration supports a local backbone path.

llama.cpp

The .gguf file is a standard llama.cpp-compatible speech-LM checkpoint, but end-to-end TTS still requires NeuCodec and the neutts preprocessing pipeline. For production use, prefer the official neutts package rather than raw llama-cli on the backbone alone.

Limitations

• Backbone only in this repo — you must supply NeuCodec and tokenization via neutts.
• Quality depends on reference cleanliness (mono, ~3–15 s, low noise, matching transcript).
• Multilingual mixing across Nano language checkpoints is not supported; pick the matching language repo.
• Upstream models may apply Perth watermarking by default; see Neuphonic documentation.
• Benchmarks in upstream READMEs measure LM token throughput, not full codec-inclusive latency.

License

Follows the NeuTTS Open License 1.0 (same family as upstream Neuphonic Nano releases). See the license text in neuphonic/neutts-nano / neuphonic/neutts-nano-q8-gguf (LICENCE file). Use only in compliance with upstream terms and applicable law.

Citation / attribution

Created by Neuphonic. This aoiandroid repo hosts mirrored Q8 weights for app distribution; credit the upstream authors when publishing work that uses these models.