An open-source text-to-speech model that packs a complete English text-to-waveform pipeline into roughly 9.4 million parameters — full local speech synthesis in a ~38 MB file that runs on a plain CPU.
Last verified · 2026-07-26 · by Moe Ameen
Inflect-Micro-v2 is a compact text-to-speech (TTS) model released on Hugging Face by Owen Song, an independent developer, under the permissive Apache-2.0 license. Its whole pitch is size: it packs a complete local text-to-waveform pipeline — the English phoneme frontend, the acoustic model, and the vocoder that produces the actual audio — into about 9.36 million parameters, a roughly 37.5 MB file in FP32. "Full speech capabilities" here means it goes straight from typed English to a finished 24 kHz mono waveform on its own, with no separate vocoder or cloud call in the loop.
Architecturally it sits in the VITS family of end-to-end TTS systems, so the text-to-audio step is a single model rather than a chain of components you have to wire together. It runs on a plain CPU or on CUDA; the model card reports throughput around 6.28× real time on a 4-thread CPU, meaning a minute of speech renders in roughly ten seconds without a GPU. Output is deterministic when you fix the seed, and you get a handful of controls — speaking speed (about 0.5×–2.0×) and a variation dial (0.0–1.0) — plus punctuation-aware segmentation so longer passages stay stable.
Be clear about the scope, because it is deliberately narrow. Inflect-Micro-v2 speaks English only, in a single fixed male voice. It does not do voice cloning, multiple speakers, or other languages — the author frames those as explicit non-goals in service of the tiny footprint. On community blind listening it landed at a 66.2% human-preference rate with a UTMOS22 naturalness score of about 4.4 and a two-ASR semantic word error rate near 4%, which is genuinely strong for a sub-10M-parameter model but still short of large cloud voices. It's the quality-leaning member of the "Inflect v2" pair; a smaller sibling, Inflect-Nano-v2, prioritizes an even smaller footprint under 4M parameters.
Treat the specifics above as a snapshot from the model card and its community discussion, and confirm current numbers on the Hugging Face page before quoting them.
Inflect-Micro-v2 lives at the very start of the pipeline — the raw "turn this sentence into sound" step — and it does that one thing in a 38 MB file with no cloud dependency. That makes it a great script-audition and scratch-voiceover tool sitting upstream of finished content. The natural workflow: draft your hook and script, run it through Inflect-Micro-v2 locally to hear the pacing and catch clunky lines before you invest in production, tighten the copy, and then bring the finished idea into Kompozy — where the shipping content actually gets made. Kompozy generates the polished spoken voice itself through HeyGen's native TTS on Persona Shorts and Persona HeyGen avatar video, burns in word-synced auto-captions, adds optional B-roll or a VFX hook, and reframes to 9:16, 1:1, and 16:9.
The division of labor is honest: Inflect-Micro-v2 is a free, offline, single-voice narration engine for prototyping and low-cost bulk audio; Kompozy is the generation-and-publishing engine that turns a rough script into a full week of on-brand output — talking-head video, Carousels, Photo Posts, Quote Graphics, a Blog Article, an Email Newsletter, and native Text Posts — all governed by a Persona Brief and fanned across Instagram, TikTok, YouTube, LinkedIn, Facebook, X, Pinterest, and Threads plus blog and email from one queue. Use the little model to pressure-test the words; use Kompozy to ship the finished, captioned, multi-platform content those words become.
It is an open-source (Apache-2.0) text-to-speech model by independent developer Owen Song that fits a complete English text-to-waveform pipeline into about 9.36 million parameters — roughly a 37.5 MB file that runs offline on a CPU and outputs 24 kHz mono audio in a single fixed male voice.
No. It is deliberately narrow: English only, one fixed male voice, no voice cloning and no multi-speaker support. The author treats those as non-goals to keep the model under 10M parameters. If you need cloning or many voices, a larger cloud TTS is the right tool.
Strong for its size but not cloud-grade. Community blind testing put it around a 66.2% human-preference rate with a UTMOS22 naturalness score near 4.4 and a two-ASR word error rate around 4%. Listeners describe it as clearly non-robotic, though inflections can flatten on unfamiliar phrasing.
Kompozy generates its own spoken voice through HeyGen native TTS on avatar-video formats rather than importing external WAV files, so the practical pairing is workflow-based: use Inflect-Micro-v2 to audition and scratch-voice your script locally, then generate the finished, captioned, multi-platform content in Kompozy.
Very little. It runs on a plain CPU or on CUDA, and the model card reports about 6.28× real-time throughput on a 4-thread CPU — so a minute of speech renders in roughly ten seconds without a GPU, which is why it suits offline, embedded, and low-resource use.