neutts

NeuTTS

HuggingFace :

• NeuTTS-Air (English): Model, Q8 GGUF, Q4 GGUF, Space

• NeuTTS-Nano Multilingual Collection:

  • NeuTTS-Nano (English): Model, Q8 GGUF, Q4 GGUF
  • NeuTTS-Nano-French: Model, Q8 GGUF, Q4 GGUF
  • NeuTTS-Nano-German: Model, Q8 GGUF, Q4 GGUF
  • NeuTTS-Nano-Spanish: Model, Q8 GGUF, Q4 GGUF
  • Multilingual Space

• NeuTTS-2E (English, emotional control): Model, Q8 GGUF, Q4 GGUF, Space

NeuTTS-2E Demo Video

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

State-of-the-art Voice AI has been locked behind web APIs for too long. NeuTTS is a collection of open source, on-device, TTS speech language models with instant voice cloning. Built off of LLM backbones, NeuTTS brings natural-sounding speech, real-time performance, built-in security and speaker cloning to your local device - unlocking a new category of embedded voice agents, assistants, toys, and compliance-safe apps.

Key Features

• Best-in-class realism for their size - produce natural, ultra-realistic voices that sound human, at the sweet spot between speed, size, and quality for real-world applications
• Optimised for on-device deployment - quantisations provided in GGUF format, ready to run on phones, laptops, or even Raspberry Pis
• Instant voice cloning - create your own speaker with as little as 3 seconds of audio
• Simple LM + codec architecture - making development and deployment simple

[!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 models are built from small LLM backbones - lightweight yet capable language models optimised for text understanding and generation - as well as a powerful combination of technologies designed for efficiency and quality:

• Supported Languages: English, Spanish, German, French (model-dependent)
• Audio Codec: NeuCodec - our 50hz neural audio codec that achieves exceptional audio quality at low bitrates using a single codebook
• Context Window: 2048 tokens, enough for processing ~30 seconds of audio (including prompt duration)
• Format: Quantisations available in GGUF format for efficient on-device inference
• Responsibility: Watermarked outputs
• Inference Speed: Real-time generation on mid-range devices
• Power Consumption: Optimised for mobile and embedded devices

	NeuTTS-Air	NeuTTS-Nano Models	NeuTTS-2E
# Params (Active)	~360m	~120m	~125m
# Params (Emb + Active)	~552m	~229m	~236m
Input Format	Phonemes	Phonemes	Text
License	Apache 2.0	NeuTTS Open License 1.0	NeuTTS Open License 1.0

Backbone Variants

Model	PyTorch repo	GGUF repos	Cloning	Emotions	Streaming
NeuTTS-Air	neuphonic/neutts-air	neuphonic/neutts-air-{q4,q8}-gguf	Yes	—	GGUF only
NeuTTS-Nano	neuphonic/neutts-nano	neuphonic/neutts-nano-{q4,q8}-gguf	Yes	—	GGUF only
NeuTTS-Nano German	neuphonic/neutts-nano-german	neuphonic/neutts-nano-german-{q4,q8}-gguf	Yes	—	GGUF only
NeuTTS-Nano French	neuphonic/neutts-nano-french	neuphonic/neutts-nano-french-{q4,q8}-gguf	Yes	—	GGUF only
NeuTTS-Nano Spanish	neuphonic/neutts-nano-spanish	neuphonic/neutts-nano-spanish-{q4,q8}-gguf	Yes	—	GGUF only
NeuTTS-2E	neuphonic/neutts-2e	neuphonic/neutts-2e-{q4,q8}-gguf	4 fixed speakers	6 + neutral	GGUF only

Codec Variants

Codec	Engine	Encode	Decode	GPU support
neuphonic/neucodec	PyTorch	Yes	Yes	Yes
neuphonic/distill-neucodec	PyTorch	Yes	Yes	Yes
neuphonic/neucodec-onnx-decoder	ONNX	—	Yes	—
neuphonic/neucodec-onnx-decoder-int8	ONNX	—	Yes	—

Any backbone can be paired with any codec. Decoder-only (ONNX) codecs require pre-encoded references — these ship with the repo for all bundled voices, including the NeuTTS-2E speakers.

Throughput Benchmarking

These benchmarks are for the Q4_0 quantisations neutts-air-Q4_0 and neutts-nano-Q4_0. Note that all models in the NeuTTS-Nano Multilingual Collection have an identical architecture, so these results should apply for any Q4_0 model in the collection.

CPU benchmarking used llama-bench (from llama.cpp) to measure prefill and decode throughput at multiple context sizes. For the GPU benchmark (RTX 4090), we leverage vLLM to maximise throughput, using the vLLM benchmark.

We include benchmarks on four devices: Galaxy A25 5G, AMD Ryzen 9HX 370, iMac M4 16GB, NVIDIA GeForce RTX 4090.

	NeuTTS-Air	NeuTTS-Nano
Galaxy A25 5G (CPU only)	20 tokens/s	45 tokens/s
AMD Ryzen 9 HX 370 (CPU only)	119 tokens/s	221 tokens/s
iMAc M4 16 GB (CPU only)	111 tokens/s	195 tokens/s
RTX 4090	16194 tokens/s	19268 tokens/s

[!NOTE] llama-bench used 14 threads for prefill and 16 threads for decode (as configured in the benchmark run) on AMD Ryzen 9HX 370 and iMac M4 16GB, and 6 threads for each on the Galaxy A25 5G. The tokens/s reported are when having 500 prefill tokens and generating 250 output tokens.

[!NOTE] Please note that these benchmarks only include the Speech Language Model and do not include the Codec which is needed for a full audio generation pipeline.

Get Started with NeuTTS

[!NOTE] We have added a streaming example using the llama-cpp-python library as well as a finetuning script. For finetuning, please refer to the finetune guide for more details.

1. Install NeuTTS

   pip install neutts
   

   Or for a local editable install, clone this 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.

   To use any of the GGUF backbones (e.g., in basic_streaming_example.py) you need to install the llama-cpp-python package.

   For the best performance, you must compile this package from source with hardware acceleration enabled for your specific operating system and target device (CPU or GPU).

   macOS (Apple Silicon)

   For M-series Macs, it is highly recommended to use Apple's native Accelerate framework for optimized CPU performance:

      CMAKE_ARGS="-DGGML_METAL=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=Apple" pip install "neutts[llama]" --force-reinstall --no-cache-dir
   

   Linux (OpenBLAS)

   For Linux, you can accelerate CPU performance using OpenBLAS.

   Prerequisite: Ensure you have OpenBLAS installed on your system (e.g., sudo apt-get install libopenblas-dev on Ubuntu). For other distros, refer to the OpenBLAS Installation Guide.

      CMAKE_ARGS="-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS" pip install "neutts[llama]" --force-reinstall --no-cache-dir
   

   Windows (OpenBLAS)

   Prerequisite: Ensure you have OpenBLAS installed on your system. Please refer to the OpenBLAS Installation Guide.

   For Windows users utilizing PowerShell, set the environment variable and run the install command like this:

      $env:CMAKE_ARGS="-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS"; pip install "neutts[llama]" --force-reinstall --no-cache-dir
   

   Looking for GPU Support?

   If you have a dedicated GPU (Nvidia/CUDA, AMD/ROCm, M-Series Mac/Metal) and want to utilize it instead of the CPU, the CMAKE flags will be different.Please refer to the official llama-cpp-python documentation for the exact flags required for your specific hardware.

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

   pip install "neutts[onnx]"
   

Examples

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

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

Several examples are available, including a Jupyter notebook in the examples folder.

Basic Example

Run the basic example script to synthesize speech:

python -m examples.basic_example \
  --input_text "My name is Andy. I'm 25 and I just moved to London. The underground is pretty confusing, but it gets me around in no time at all." \
  --ref_audio samples/jo.wav \
  --ref_text samples/jo.txt

To specify a particular model repo for the backbone or codec, add the --backbone and --codec arguments. Available backbones are listed in the NeuTTS-Air, NeuTTS-Nano Multilingual Collection and NeuTTS-2E huggingface collections.

[!CAUTION] If you are using a non-English backbone, it is highly recommended to use a same-language reference for best performance. See the 'example reference files' section below to select an appropriate example reference.

One-Code Block Usage

…

Streaming

Speech can also be synthesised in streaming mode, where audio is generated in chunks and plays as generated. The streaming example scripts use pyaudio for playback, which needs the PortAudio library on macOS and Linux:

# macOS
brew install portaudio && pip install pyaudio

# Ubuntu/Debian
sudo apt install portaudio19-dev && pip install pyaudio

# Windows (prebuilt wheels)
pip install pyaudio

Then run:

python -m examples.basic_streaming_example \
  --input_text "My name is Andy. I'm 25 and I just moved to London. The underground is pretty confusing, but it gets me around in no time at all." \
  --ref_codes samples/jo.pt \
  --ref_text samples/jo.txt

Again, a particular model repo can be s