NeuTTS-2E

🚀 Spaces Demo, 🔧 Github

Q8 GGUF version, Q4 GGUF version

Created by Neuphonic - building faster, smaller, on-device voice AI

NeuTTS-2E is a super-fast, highly realistic, on-device emotional TTS speech language model. It is an early alpha release, English-only model, supporting six emotions plus neutral (angry, disgusted, fearful, happy, sad, surprised and neutral) across four fixed speakers (emily, paul, sophie, steven). With a compact backbone and an efficient LM + codec design, NeuTTS-2E delivers strong naturalness and expressive control at a fraction of the compute, making it ideal for embedded voice agents, assistants, toys, and privacy-sensitive applications.

[!NOTE] This model is English only with fixed speakers: for other languages and instant voice cloning, see the NeuTTS Nano Multilingual Collection.

Key Features

• ⚡️ Ultra-fast for on-device — built for real-time or better-than-real-time generation on laptop-class CPUs
• 😠😁😭 Emotional control — six emotions plus neutral, selected with a single argument
• 🗣 High realism for its size — natural, expressive speech in a compact footprint
• 📦 GGUF/GGML-friendly deployment — easy to run locally via CPU-first tooling
• 🔒 Local-first + compliance-friendly — keep audio and text on-device

[!CAUTION] Websites like neutts.com are popping up and they're not affliated with Neuphonic, our github or this repo.

We are on neuphonic.com only. Please be careful out there! 🙏

Model Details

NeuTTS-2E is designed for maximum speed per parameter while retaining strong naturalness and expressive control:

• Backbone: compact LM backbone tuned for emotional TTS token generation
• Input Format: text — no phonemizer or system dependencies required
• Speakers: four fixed speakers (emily, paul, sophie, steven)
• Emotions: angry, disgusted, fearful, happy, sad, surprised and neutral
• Audio Codec: NeuCodec - our open-source neural audio codec that achieves exceptional audio quality at low bitrates using a single codebook
• Format: quantisations available in GGUF format for efficient on-device inference
• Responsibility: Watermarked outputs
• Inference Speed: Optimised for real-time generation on CPUs
• Power Consumption: Designed for mobile and embedded devices

Parameter Count

• Active params (backbone only): ~125M
• Total params (backbone + tied embeddings/head): ~236M

Get Started with NeuTTS

1. Install NeuTTS

   pip install neutts
   

   Or for a local editable install, clone the neutts repository and run in the base folder:

   pip install -e .
   

   Alternatively to install all dependencies, including onnxruntime and llama-cpp-python (equivalent to steps 2 and 3 below):

   pip install neutts[all]
   

   or for an editable install:

   pip install -e .[all]
   

2. (Optional) Install llama-cpp-python to use .gguf models.

   pip install "neutts[llama]"
   

   Note that this installs llama-cpp-python without GPU support. To install with GPU support (e.g., CUDA, MPS) please refer to: https://pypi.org/project/llama-cpp-python/

3. (Optional) Install onnxruntime to use the .onnx decoder.

   pip install "neutts[onnx]"
   

Examples

To get started with the example scripts, clone the neutts repository and navigate into the project directory:

git clone https://github.com/neuphonic/neutts.git
cd neutts

Basic Example

Run the emotional example script to synthesize speech:

python -m examples.basic_example_emotions \
  --input_text "I can't believe it's finally here!" \
  --speaker emily \
  --emotion happy

Simple One-Code Block Usage

from neutts import NeuTTS2E
import soundfile as sf

tts = NeuTTS2E()

wav = tts.infer(
    "I can't believe it's finally here!",
    speaker="emily",
    emotion="happy",
)
sf.write("test.wav", wav, 24000)

Responsibility

Every audio file generated by NeuTTS-2E includes by default a Perth (Perceptual Threshold) Watermark.

Disclaimer

Don't use this model to do bad things… please.