Inflect-Nano brings TTS to potato PCs with just 4.63M parameters

Owensong has released Inflect-Nano-v1, an ultra-compact text-to-speech model weighing just 4.63 million total inference parameters (3.46M acoustic model + 1.17M vocoder). This makes it one of the smallest usable neural TTS models available, capable of running on extremely low-spec hardware—even what the developer calls a "potato computer." The model produces 24kHz audio, currently supports English with a single male voice, and runs locally via a simple PyTorch inference script. For context, Inflect-Nano is roughly 17x smaller than Kokoro, 108x smaller than Chatterbox, and nearly 1000x smaller than Fish Audio S2 Pro.

Despite its impressive size-to-functionality ratio, the model has clear limitations: speech can sound robotic, it struggles with difficult or unseen text, and the lightweight vocoder remains a bottleneck for quality. Nonetheless, it serves as an interesting baseline for extremely tiny local speech synthesis, targeted at use cases like offline assistants, embedded devices, browser/WASM-style projects, and local voice agents. The developer is open to releasing a v2 with a larger training budget if the community finds it useful. The model is available on Hugging Face with audio examples.