NeuTTS Air: open source voice model that operates on devices without internet

Neuphonic has launched NeuTTS Air as an open source project, a speech language model that runs directly on local devices. Unlike advanced speech synthesis systems that have traditionally been available only through cloud APIs, this model operates completely without internet connection. The company indicates this is the first speech synthesis model with these realistic characteristics capable of running entirely on the user's own device.

The model is built on Qwen 0.5B, a lightweight language model optimized for text understanding and generation, combined with NeuCodec, Neuphonic's proprietary neural audio codec. This architecture allows the system to run in real-time even on mid-range devices, including laptops, mobile phones, and Raspberry Pi boards. The company has distributed NeuTTS Air in GGML format, designed specifically to enable efficient inference on devices without requiring specialized hardware.

One of the system's standout features is instant voice cloning, which allows creating a personalized voice profile with just three seconds of reference audio. This functionality operates entirely on the local device, meaning voice data is never transmitted to external servers. Neuphonic notes that this approach addresses privacy and regulatory compliance concerns, especially relevant in applications handling sensitive data.

The model generates voices with a high degree of naturalness for its size, balancing audio quality with processing speed and storage requirements. The architecture combines a compact language model with an audio codec that achieves high quality at reduced bitrates through the use of a single codebook. According to the company, this balance enables real-time applications on devices with limited resources.

NeuTTS Air is freely available on Hugging Face under an open source license. The model's audio outputs include watermarks for identification purposes. Neuphonic indicates that the system's power consumption has been specifically optimized for mobile and embedded devices, enabling its use in applications ranging from voice assistants to interactive toys and tools requiring strict privacy regulation compliance.

Key points

• NeuTTS Air operates entirely on local devices without internet connection.
• Voice data remains on the device itself to ensure privacy and regulatory compliance.
• Released as free open source on Hugging Face.
• The system enables instant voice cloning with just three seconds of reference audio.
• Built on Qwen 0.5B and uses NeuCodec, Neuphonic's proprietary neural audio codec.
• Runs in real-time on mid-range devices, including laptops, mobile phones, and Raspberry Pi.
• Audio outputs include watermarks for identification.