Inflect-Nano, a 4.63M-parameter local TTS model with its own vocoder (huggingface.co)

Inflect-Nano, a new compact text-to-speech (TTS) model with 4.63 million parameters, has gained traction as the third most popular TTS model on Hugging Face's leaderboard. Unlike larger TTS models, Inflect-Nano is designed for local deployment and aims to explore the limits of ultra-lightweight speech synthesis. The model integrates its own vocoder, enabling a complete text-to-waveform stack that can produce 24 kHz audio with a single male English voice, making it suitable for small-scale applications like offline assistants, embedded demos, and efficient inference research. This innovation is noteworthy for the AI/ML community as it highlights the potential for minimal TTS systems to deliver viable performance for specific tasks, particularly in environments with limited resources. Inflect-Nano employs a non-autoregressive FastSpeech-style acoustic model and a compact HiFi-GAN-style vocoder to achieve its efficiency. However, the model's experimental nature means that while it showcases capabilities in real-time applications, users should be aware of its limitations, such as sounding robotic and struggling with complex text. This release paves the way for further research in sub-5M TTS models, potentially broadening the scope of accessible speech synthesis technology.