AIArticle Inflect-Micro-v2 isn't the breakthrough its headline implies, but it resets the floor for offline speech on cheap CPUs.
Rachel Goldstein A solo developer named Owen Song just shipped Inflect-Micro-v2, an English text-to-speech model that goes from raw text to a 24 kHz waveform in 9,356,513 parameters — a 37.5 MB FP32 file, Apache-2.0 licensed, running 6.28× faster than real time on four CPU threads. It hit the Hacker News front page under the headline "complete voice in 9.36M parameters," and the first thing to say is that the headline oversells it: this is TTS only, one fixed synthetic male voice, English only, no cloning. More than one commenter showed up hoping "complete voice" meant speech recognition too. It doesn't.
The second thing to say is that once you strip the framing, what's left is still worth your attention — just not for the reason the headline implies.
The recipe is old; the compression is the story #
Sub-100M-parameter local TTS is a well-trodden path. Piper has been running VITS-family voices on Raspberry Pis for years and quietly became the default speech engine for the Home Assistant crowd. KittenTTS made the same HN splash in 2025 with a 15M-parameter nano model under 25 MB, and its family scales up to 80M. And Kokoro-82M proved that 82M parameters gets you quality good enough to top community arena rankings. Inflect-Micro-v2 doesn't introduce a new architecture — it's a VITS-style stack: eSpeak-ng phoneme frontend, monotonic alignment, a small flow, and an integrated neural waveform decoder.
That word "integrated" is doing the honest work the word "complete" was trying to do. Most tiny TTS pipelines are two models pretending to be one: an acoustic model that predicts spectrograms, plus a separate vocoder that turns them into audio, each with its own weights, export path, and failure modes. Inflect-Micro-v2 is a single text-to-waveform network, and the 9.36M figure covers everything you deploy, vocoder included. Song's companion release, Inflect-Nano-v1, pushes the same idea under 5M parameters — while candidly describing itself as robotic and research-only, which tells you 9–10M is roughly where this recipe stops sounding like a toy.
So the real headline is a trend line, not a breakthrough: 82M (Kokoro, quality-competitive) → 15M (KittenTTS nano, usable) → 9.36M (Inflect, usable with caveats). The floor for "acceptable offline English voice" has dropped by roughly an order of magnitude in about eighteen months, on a 2021-era architecture, without any of these projects coming from a lab with a GPU cluster. Song reports spending a little over US$500 of his own money on training and evaluation.
Read the eval like an adversary #
The model card is unusually disciplined for a hobbyist release: a frozen protocol of 400 unseen prompts, a UTMOS22 naturalness score of 4.395, 3.99% semantic word error rate under a two-model ASR consensus, and a 66.2% preference rate from a community listening study. Deterministic seeds mean you can reproduce an exact utterance — genuinely rare in TTS releases and very useful for regression-testing a voice pipeline.
But hold the numbers to the light. UTMOS is an automatic MOS predictor, and small VITS models are exactly the distribution it was trained to score — a 4.4 from a machine judge is not a 4.4 from humans. The preference study is 21 wins, 10 losses, and 3 ties: thirty-four votes. And the human signal from HN was split in a revealing way. Listeners agreed the quality is startling for the size, then disagreed about whether it's pleasant in absolute terms — one compared it favorably to anything from a decade ago, another put it "marginally better" than macOS's ancient Samantha voice. Both can be true. This is a voice you'd happily accept from a $15 device and quickly tire of in a two-hour audiobook.
Where 9M parameters actually matters #
The practical question isn't "is this the best local TTS" — it isn't; Kokoro-82M and Piper's high-quality voices beat it on naturalness and voice selection. The question is what changes when decent speech costs 37 MB and a fraction of one CPU core.
The answer showed up in the HN thread within hours: someone had already wired the ONNX export into speech-dispatcher, the Linux accessibility layer, replacing their old model. That's the adoption pattern to expect — screen readers, kiosk and appliance UIs, offline voice feedback in CLI tools and games, single-board-computer assistants where Piper's larger voices strain the thermal budget. At 6.28× real time on four threads, you can synthesize speech as an afterthought inside a process that's busy doing something else. Getting started is genuinely trivial: hf download owensong/Inflect-Micro-v2
, point the bundled InflectTTS
class at the folder, and you're writing WAVs on CPU, with speed and variation knobs and no reference audio required. Two trade-offs deserve more attention than the model card gives them. First, one voice and one language is a product constraint, not a demo footnote — if your app needs a female voice, a second language, or brand differentiation, you're back to Piper or Kokoro regardless of size. (The parameter count cuts the other way too: several HN commenters noted a model this small should be cheap to fine-tune, and an Apache-2.0 license invites exactly that. Expect community voices if the project gets traction.) Second, the Apache license covers the weights, but the phoneme frontend is eSpeak-ng, which is GPL — the same licensing wrinkle that pushed Piper's maintainers to relicense their rewrite. If you're shipping a proprietary embedded product, your lawyers will care about the text frontend long before they care about the flow blocks.
The verdict #
Inflect-Micro-v2 is not a capability breakthrough, and "complete voice" is the wrong frame. It's something more interesting for where the field is heading: evidence that competent, honestly-evaluated, permissively-licensed speech synthesis is now a solo-developer, sub-$1,000 project. The moat around basic TTS is gone — what's left to differentiate on is voices, languages, expressiveness, and streaming latency, which is exactly where the 80M+ models and the cloud APIs still earn their keep. If you maintain anything that speaks from a CPU-bound or offline environment, this is worth an afternoon's evaluation. If you were waiting for tiny models to make ElevenLabs nervous, keep waiting — but notice how fast the floor is rising toward the ceiling.
Sources & further reading #
[Inflect-Micro-v2 model card](https://huggingface.co/owensong/Inflect-Micro-v2)— huggingface.co -
[Inflect-Micro-v2: complete voice in 9.36M parameters](https://news.ycombinator.com/item?id=49053375)— news.ycombinator.com -
[Inflect-Nano-v1 model card](https://huggingface.co/owensong/Inflect-Nano-v1)— huggingface.co -
[KittenTTS: TTS model under 25MB](https://github.com/KittenML/KittenTTS)— github.com -
[Piper: fast local neural TTS](https://github.com/rhasspy/piper)— github.com
[Rachel Goldstein](https://sourcefeed.dev/u/rachel_goldstein)· Dev Tools Editor
Rachel has been embedded in the developer tooling ecosystem for nearly eight years, covering everything from IDE wars and package-manager drama to the quiet rise of AI-assisted coding. She has a soft spot for open-source maintainers and an unhealthy number of terminal emulators installed on a single laptop.
Discussion 0 #
No comments yet
Be the first to weigh in.