Inflect-Micro-v2 Released: Text-to-Speech Model Under 10M Parameters

Introduction of Inflect-Micro-v2

Inflect-Micro-v2 is a new text-to-waveform speech synthesis model designed for local deployment. It operates with 9,356,513 deployable parameters, totaling 37.53 MB in FP32, and produces 24 kHz mono output. The model focuses on providing complete English text-to-speech capabilities within a small footprint.

Key Features and Capabilities

The model offers fixed-voice English TTS, deterministic seeds for consistent output, and robust handling of long texts. It supports both CPU and CUDA inference, making it versatile for different computing environments. Inflect v2 also includes a smaller variant, Inflect-Nano, which prioritizes an even smaller footprint (under 4M parameters) while Micro focuses on quality under 10M parameters.

Performance and Evaluation

Inflect-Micro-v2 was evaluated against established compact TTS baselines such as KittenTTS Nano, Piper Low, and Supertonic 3. In an anonymous community study, Inflect-Micro-v2 recorded a 66.2% preference rate. The model also reported a UTMOS22 score of 4.395, with a 95% bootstrap CI of 4.381–4.408. The evaluation methodology avoids single-metric scores, instead reporting human preference, predicted naturalness, multi-ASR intelligibility, footprint, and runtime separately.

Implications for Local TTS

The release of Inflect-Micro-v2 demonstrates progress in developing highly compact text-to-speech models. Its small parameter count allows for efficient local deployment, which could be beneficial for applications requiring on-device speech synthesis without extensive computational resources or cloud connectivity. The project is independently funded, with potential for future expansion based on community interest.