Neuphonic Open-Sources NeuTTS Air: A 748M-Parameter On-Device Speech Language Model with Instant Voice Cloning

Neuphonic has introduced NeuTTS Air, an innovative open-source text-to-speech (TTS) speech language model engineered for seamless, real-time operation on CPUs without relying on cloud services. This model features 748 million parameters based on the Qwen2 architecture and is distributed in GGUF quantized formats (Q4/Q8), enabling efficient inference through llama.cpp or llama-cpp-python. Licensed under Apache-2.0, NeuTTS Air comes with ready-to-run demos and comprehensive examples to facilitate immediate use.

Introducing NeuTTS Air: A New Era in On-Device TTS

NeuTTS Air combines a 0.5-billion parameter Qwen backbone with Neuphonic’s proprietary NeuCodec audio codec, delivering a “super-realistic” TTS experience directly on user devices. The system excels at cloning voices from as little as three seconds of reference audio, synthesizing speech that closely mimics the original speaker’s style. This makes it ideal for voice assistants and applications where privacy is paramount. The model’s documentation highlights its capability for real-time CPU-based synthesis and compact deployment footprints.

Core Advantages of NeuTTS Air

• High-fidelity speech at a compact scale: Achieves natural prosody and voice timbre preservation with a lightweight ~0.7B parameter Qwen2-class TTS model.
• Optimized for local devices: Distributed in GGUF format with Q4 and Q8 quantizations, enabling smooth performance on laptops, smartphones, and single-board computers like Raspberry Pi.
• Rapid voice cloning: Transfers vocal style from approximately 3 seconds of clean reference audio paired with its transcript.
• Efficient model and codec integration: Combines the Qwen 0.5B backbone with NeuCodec operating at 0.8 kbps and 24 kHz sampling rate, balancing latency, model size, and audio quality.

Technical Composition and Execution Workflow

• Model backbone: Utilizes the Qwen 0.5B language model as a lightweight core for conditioning speech generation. The hosted model contains 748 million parameters under the qwen2 architecture.
• Audio codec: NeuCodec compresses and decompresses acoustic tokens at a low bitrate of 0.8 kbps with a high-fidelity output of 24 kHz, enabling compact and efficient audio representation.
• Quantization and compatibility: Pre-quantized GGUF models (Q4 and Q8) are provided, with detailed instructions for running inference using llama-cpp-python and an optional ONNX decoder for enhanced flexibility.
• Dependencies and tools: Employs espeak for phoneme conversion, accompanied by example scripts and a Jupyter notebook to demonstrate end-to-end speech synthesis.

Performance Tailored for On-Device Use

NeuTTS Air is engineered to deliver real-time speech synthesis on mid-tier hardware, prioritizing CPU-first execution. Its GGUF quantized models are optimized for devices ranging from standard laptops to compact single-board computers. Although specific frame rates or real-time factors (RTF) are not disclosed, the included demos and hosted Spaces confirm its capability for local inference without GPU acceleration.

Voice Cloning Process Explained

The voice cloning workflow requires two inputs: (1) a reference WAV audio file and (2) the corresponding transcript text. NeuTTS Air encodes the reference audio into style tokens, which it then uses to generate speech in the same vocal timbre for any arbitrary text. The recommended reference audio length is between 3 to 15 seconds of clean, mono sound. Pre-encoded sample voices are also provided to assist users in testing.

Privacy, Ethical Use, and Audio Watermarking

Designed with privacy in mind, NeuTTS Air ensures that all audio and text data remain on the user’s device unless explicitly shared. To promote responsible deployment, every generated audio clip incorporates a Perth (Perceptual Threshold) watermark, which helps verify authenticity and traceability of synthesized speech.

Positioning Among Local TTS Solutions

While several open-source, local TTS frameworks exist-many leveraging GGUF quantized models-NeuTTS Air distinguishes itself by integrating a compact language model with a neural codec, enabling instant voice cloning, CPU-optimized quantizations, and embedded watermarking under a permissive Apache-2.0 license. The claim of being the “world’s first super-realistic, on-device speech LM” is supported by its unique combination of model size, format, cloning speed, licensing, and ready-to-use runtime environments.

Insights and Future Directions

NeuTTS Air strikes a practical balance between model complexity and performance, pairing a ~0.7B parameter Qwen-class backbone with NeuCodec’s efficient 0.8 kbps codec to enable real-time, CPU-only TTS that faithfully preserves speaker characteristics from brief audio samples. Its Apache-2.0 license and built-in watermarking facilitate broad deployment, especially in privacy-sensitive contexts. To further enhance adoption and benchmarking, publishing detailed latency metrics on common CPUs and analyzing cloning quality relative to reference audio duration would be valuable. The minimal dependency footprint-relying mainly on eSpeak and llama.cpp or ONNX-also reduces privacy risks, making it well-suited for edge devices requiring offline speech synthesis without compromising intelligibility.