Owen Song releases complete local speech synthesis in 9.36M parameters (runtimewire.com)

On July 24, 2026, Owen Song unveiled Inflect-Micro-v2, a 9.36M-parameter text-to-speech model capable of generating 24 kHz waveforms entirely offline, without reliance on external services like a neural vocoder. Designed by Song as a solo developer focused on machine learning efficiency, this model aims to push the boundaries of compact yet effective speech synthesis, covering a complete TTS pipeline within its parameter count. Unlike some smaller models that omit components such as vocoders in size comparisons, Inflect-Micro-v2 integrates text encoding, duration prediction, speech generation, and waveform decoding, thus offering a self-contained synthesis solution. The significance of this release lies in its potential for localized, efficient deployment in applications requiring deterministic output without the overhead of large models or multiple voices. Inflect-Micro produces a single synthetic English male voice and supports CPU or CUDA inference, making it suitable for embedded systems or offline applications. With no plans for speed optimizations that would sacrifice output quality, Song seeks feedback from users to refine future iterations. As demand dictates, the project may evolve to incorporate additional voices and language capabilities, marking an important stride in the ongoing trend towards more efficient, localized AI model deployment within the AI/ML community.