Introduction

Inflect v3 from independent developer Owen Song is now officially released as open weights under Apache 2.0. It's a family of three local English text-to-speech models: Mini (34.31M-parameter core), Micro (11.79M) and Atom (157,190 parameters). All three share six preset voices and output native 24 kHz audio. You can run them today on your own CPU, or try them in the free Hugging Face playground first. Mini is the one with extras: eight vocal-event tags such as [laugh] and [sigh], a reference-based Voice Tuner, and an optional 48 kHz enhancer.

The v3 repos first appeared on Hugging Face as a gated "pending licensing review" preview in October. The official Apache-2.0 model cards went live on November 11 (UTC), the eight event tags landed early on November 12, and Owen announced the release on X a few minutes later.

Sources: the Inflect Mini v3, Inflect Micro v3 and Inflect Atom v3 model cards, the Inflect v3 collection, the Inflect v3 playground Space, the GitHub repo, and

(November 12, 02:20 UTC).

What Shipped

  • Name and maker: Inflect v3, by Owen Song (Hugging Face owensong). It follows Inflect v2 Nano and Micro, which picked up 450+ likes on the Hub earlier this year.
  • Modality: English text-to-speech with six preset voices: cassidy, mark, hope, james, sophia and jasper. No reference recording is needed for preset speech.
  • License: Apache 2.0 for Inflect's own code and weights. Third-party pieces (like the enrollment encoders) keep their own licenses. Each repo is the complete inference package. The training pipeline and training data are not included.
  • Where to run it: locally with Python 3.12 (hf download the full repo, not just the weights), or in the browser via the Inflect v3 Space. There's no hosted API, and none is needed.
  • Output: native 24 kHz WAV for all three sizes. Mini can also output 48 kHz through its optional enhancer.

Three Sizes

  • Mini v3: 34,310,790-parameter core, 138.4 MB FP32 weights. This is the full-featured model, with event tags, Voice Tuner, the 48 kHz enhancer and long-text chunking. It runs on CPU or CUDA, and the CUDA path gets its own optimized runtime that uses the same weights.
  • Micro v3: 11,787,634-parameter core, 48.06 MB weight file. Speech only, with the same six voices at 24 kHz, a CPU-first CLI, and conservative number handling plus text chunking. It doesn't support event tags, Voice Tuner or the enhancer.
  • Atom v3: 157,190-parameter core, 656,498-byte (0.66 MB) weight file. Instead of a learned neural vocoder, Atom generates the waveform with DSP. The default Misaki pronunciation frontend adds about 1.57M frontend parameters on top of that, so the 157K figure covers the synthesis core, not the whole pipeline. Its bundled CLI is CPU-only, with optional cpu-fused and cpu-lut backends.

Mini's Extras

  • Eight event tags: [laugh], [sigh], [gasp], [cough], [yawn], [inhale], [exhale] and [clear_throat] ([throat_clear] also works). They work with all six voices on CPU at 24 kHz. These are learned non-speech sounds placed between speech segments. They don't change the emotion of the surrounding words, and there are no whisper, angry or excited style controls.
  • Voice Tuner: enroll a 5 to 15 second reference recording that you have permission to use, and Mini saves it as a reusable voice profile. Enrollment is officially supported on Linux CPU. Owen describes this as adaptation, not guaranteed cloning: resemblance and accent can vary.
  • 48 kHz enhancer: an optional 2.85M-parameter add-on that upsamples output without touching the speech backbone. The card says to choose by ear. It can sound brighter, but it may add a faint buzzy texture on some material.

Numbers (Self-Reported)

  • Mini WER: 0.739% average across 10,128 clips (all six voices on 1,688 texts from CV3-Eval, Seed-TTS and MiniMax English sets), transcribed with Qwen3-ASR-1.7B. That's a self-run test on a corpus used during development, not a leaderboard submission.
  • Mini predicted quality: UTMOS22Strong 4.479 and UTMOSv2 3.917 on a 180-clip panel. These are automatic predictors, not human MOS scores.
  • Micro: 0.703% average WER on the same corpus, UTMOS22Strong 4.39 and UTMOSv2 3.71.
  • Mini speed: about 162x real time (RTFx 162.1) on an RTX 5090 with the optimized CUDA runtime. On a Ryzen 9 3900X CPU with two Torch threads it runs at about 1.8x real time. Both are warm, author-run probes with 12 renders, not first-audio latency numbers.
  • For context only: on the hosted Open TTS Leaderboard snapshot the card quotes (November 10), Kokoro-82M averages 0.690% WER. Owen points out that this isn't a matched comparison with Inflect's self-run numbers.

Try It

Run Mini on CPU (Python 3.12, fresh environment, outside the downloaded folder):

python -m pip install huggingface-hub==0.36.2
hf download owensong/Inflect-Mini-v3 --local-dir inflect-mini-v3
cd inflect-mini-v3
python verify_release.py
python -m pip install -r core/requirements-cpu.txt
python run.py speech --voice cassidy --text "A small idea can become something worth hearing." --subtype PCM_16 --output ../mini-cassidy.wav

Add an event tag (after installing the events component):

python -m pip install -r extras/events/component/requirements-cpu.txt
python run.py events --voice cassidy --profile base --text "You really surprised me. [laugh] That was a good one." --output ../mini-laugh.wav

Turn on the 48 kHz enhancer:

python run.py speech --profile enhanced --voice hope --text "The next chapter starts here." --subtype PCM_16 --output ../mini-48k.wav

Micro and Atom use synthesize.py instead. Micro needs --model micro, and Atom has no model flag. Both have a --verify integrity check. One gotcha: the inflect/ folder in the GitHub repo is the older v1/v2 library. The v3 runtime lives in the Hugging Face repos.

Caveats

  • English only. Check names, numbers, unusual words and long passages before using it in production.
  • Event tags add sounds, not emotion. If you need laughing-while-talking or angry delivery, this isn't that.
  • Voice Tuner's supported setup is Linux CPU. Its enrollment encoders are separate downloads with their own licenses.
  • All the benchmarks come from the author, not an independent test. Listen to the samples on each card before you commit.

Conclusion

Inflect v3 gives you one voice set in three sizes. Mini is the 34M all-rounder with laughs, sighs, reference tuning and 48 kHz output. Micro is a 12M speech core for compact apps. Atom fits its synthesis core in 0.66 MB and drives a DSP waveform generator on CPU. All three are Apache 2.0 and run on your own machine. Start with the playground Space, then pull Mini if you want the extras or Micro if you want the smallest neural option.

—Anabel ♡