Small, reproducible TTS matters because many real-world apps need believable speech without cloud calls or large footprints. Inflect-Micro-v2 targets that gap: a complete local text-to-waveform pipeline in ~9.36M parameters (37.53 MB FP32) with repeatable sampling, measured quality, and practical CPU throughput.
Key Capabilities
- Compact local inference: 9,356,513 parameters; complete FP32 package ~37.53 MB; runs on CPU or CUDA with an official ONNX export for torch-free deployment.
- Repeatability and long-text support: fixed seeds reproduce the same latent sample on the same runtime stack; long inputs are split at punctuation-aware boundaries and stitched with controlled pauses and fades.
- Measured quality and intelligibility: community blind-preference reported at 66.2% in the release study, UTMOS22 predicted naturalness 4.395, and multi-ASR semantic WERs in the low single-digit percentages on held-out prompts.
- Practical runtime: reported steady-state CPU throughput around 6.28× real-time on a 4-thread reference run; an ONNX runtime path supports dynamic lengths and multiple providers.
Who it's for and trade-offs
Great fit if you need a deployable, auditable local English TTS with a tiny footprint—edge devices, privacy-sensitive apps, or offline tooling where a single fixed voice is acceptable. Look elsewhere if you need multi‑language support, selectable or cloned voices, large-model long-form prosody, or production-grade accessibility guarantees (medical/legal/emergency use is not validated).
Where it fits
Inflect-Micro-v2 sits among compact local TTS systems as the quality-focused small member of the Inflect family (Nano emphasizes smaller footprint). It favors reproducible, auditable evaluation and CPU-friendly deployment rather than maximal expressiveness or multilingual coverage.
How it works (brief)
Architecturally it is a parameter-efficient VITS-family end-to-end text-to-waveform generator: English phoneme frontend, monotonic alignment, stochastic latent synthesis, coupling flows, and an integrated waveform decoder. The release includes inference code, evaluation artifacts, ONNX graphs, and a documented benchmark protocol for transparent comparisons.
