Inflect Micro v3
A compact neural voice for the things you build.
Six English voices. Native 24 kHz speech. Run it locally.
A smaller speech core
Micro brings the Inflect family to an 11,787,634-parameter neural synthesis core. Give it English text and a preset voice; the included command-line interface writes a 24 kHz WAV file on your machine.
The core count and weight-file size are not the full installation size or runtime memory requirement. Pronunciation models, dependencies and fixed voice-conditioning assets are separate from the learned synthesis-core count. Download the whole repository, not just its weights.
Run Micro locally: this model repository is the complete v3 inference package, not just weights. Download and first generation | Choose a model and find the right files. The GitHub
inflect/folder is the older v1/v2 library, not the v3 runtime.
Open TTS Leaderboard: 0.73% WER (7/50) · 4.37 UTMOS22 · 71.87x real time (H200). English results and test settings
Voices
Six English voices. Native 24 kHz Micro output. Read each exact input below, or try your own text.
Cassidy
NATIVE 24 kHz
Could you bring the smaller suitcase downstairs?
Mark
NATIVE 24 kHz
Beyond the orchard, a narrow footpath curved toward the ruined tower.
Hope
NATIVE 24 kHz
Press the side button once, then wait until the indicator turns green.
James
NATIVE 24 kHz
A shadow becomes longer when the light source moves closer to the horizon.
Sophia
NATIVE 24 kHz
We moved the chairs inside and wiped the wooden table dry.
Jasper
NATIVE 24 kHz
I thought you were arriving tomorrow, but it is lovely to see you.
These are actual generated samples, not reference recordings. The same voice reads the same text across the v3 cards for model comparison. Different voices read different texts; this is not a voice-ranking test. Audio provenance.
Voice IDs: cassidy, mark, hope, james, sophia, jasper. No reference recording is needed.
Run locally
Use a fresh Python 3.12 environment outside the verified package. Keep environment files, profiles and generated audio outside the download.
Start in a Python 3.12 environment and download the complete package:
python -m pip install huggingface-hub==0.36.2
hf download owensong/Inflect-Micro-v3 --local-dir inflect-micro-v3
cd inflect-micro-v3
Install the provided CPU dependencies, verify the files, and make your first clip:
python -m pip install -r requirements-cpu.txt
python synthesize.py --verify
python synthesize.py --model micro --voice cassidy --text "A small idea can become something worth hearing." --output ../micro-cassidy.wav
Try another preset with a different output filename:
python synthesize.py --model micro --voice hope --text "A small idea can become something worth hearing." --output ../micro-hope.wav
The CLI requires --model micro. Its defaults are CPU, seed 42 and conservative
English number preparation. It validates the full input before splitting longer
text into chunks, with a short silence between chunks.
Use a new .wav path with an existing parent directory. The examples write
beside the download; existing output files are never overwritten.
Run python synthesize.py --help for the complete argument list.
Installation can require downloads. Once the dependencies and package assets are
present, speech generation runs locally. --device cuda requires a compatible
PyTorch/CUDA installation; the supplied CPU recipe is not a CUDA setup guide.
Features
| Capability | Micro |
|---|---|
| English text to speech | Included |
| Six fixed voices | Included |
| Native 24 kHz WAV output | Included |
| Conservative number preparation and text chunking | Included |
| Package-file integrity verification | python synthesize.py --verify |
Micro is a speech-only package. Event tags, the optional 48 kHz enhancer and
VoiceTuner belong to Inflect Mini v3,
not this runtime. Mini profile flags and bracketed directions such as [laugh]
are not Micro controls.
English presets do not imply support for every language, accent, symbol or input format. Unsupported input can fail explicitly; review the spoken-text metadata when using numbers or unusual notation.
Specifications
| Inside this download | Specification |
|---|---|
| Learned synthesis core | 11,787,634 parameters |
| Core weights | models/micro/weights.pt |
| Serialized core weight file | 48,056,467 bytes (48.06 MB, decimal) |
| Output | Native 24 kHz, mono PCM16 WAV |
| Voices | Six fixed English presets |
| Integrity inventory | Release manifest |
Evaluation
Open TTS Leaderboard
Micro reached 0.73% WER with roughly a third of Mini's synthesis-core parameters. These scores are from the hosted evaluation, not our local runs.
Snapshot: 18 November 2026. English, no voice cloning, averaged across Seed-TTS, CV3-Eval and MiniMax. The WER ranks are for the 50 models in that view.
| Model | Learned synthesis core | WER rank | Average WER | UTMOS22 | RTFx / hardware |
|---|---|---|---|---|---|
| Inflect Mini v3 | 34.31M | 4/50 | 0.71% | 4.49 | 48.87x / H200 |
| Inflect Micro v3 | 11.79M | 7/50 | 0.73% | 4.37 | 71.87x / H200 |
| Inflect Atom v3 | 157.19K | 19/50 | 0.85% | 3.66 | 13.17x / CPU |
| Kokoro-82M | 82M | 2/50 | 0.69% | 4.51 | 73.96x / H200 |
WER measures transcription errors (lower is better). UTMOS22 predicts speech naturalness on a 1-5 scale; it is not a human rating or UTMOSv2. RTFx is seconds of audio generated per second of compute, not time to first audio. Atom was evaluated on the host CPU; its throughput is not directly comparable with the H200 GPU results.
The hosted scores describe the default non-cloning path, not all six presets, and do not evaluate event tags, reference tuning or 48 kHz enhancement.
Live leaderboard · Snapshot release notes · Metric and evaluation code
Self-run: Micro native speech across six voices
Completed native24 base-speech results on the full 1,688-text corpus x six voices. Known development corpus; not an official submission or a new independent test. WER is percent (lower is better); quality proxies are not human MOS.
| Metric | Result | Clips |
|---|---|---|
| WER %, equal dataset then equal voice mean | 0.703185 | 10128 |
| WER %, pooled word counts (separate) | 0.843623 | 10128 |
| UTMOS22Strong, objective score | 4.392165 | 180 |
| UTMOSv2, objective score | 3.706836 | 180 |
Quality uses a separate 180-clip panel, 30 clips per voice. The self-run metrics are separate from the hosted leaderboard scores.
Evaluation and methodology / Machine-readable results / Audio provenance.
Family
Explore the official Inflect v3 collection.
Three sizes, with a different synthesis core in each package.
| Model | Focus |
|---|---|
| Inflect Mini v3 | Larger speech core and its own optional components |
| Inflect Micro v3 | Compact neural synthesis with an 11.79M-parameter core |
| Inflect Atom v3 | A 157,190-parameter synthesis core with DSP waveform generation |
Each repository includes its own runtime and dependencies. Family membership does not make model weights, flags or optional components interchangeable.
License and notices
Inflect-specific code and model weights are released under Apache-2.0. Read the license and component scope.
Upstream components retain their own licenses and notices. Keep the supplied third-party notices, Apache-2.0 text and StyleTTS2 MIT text with the components they cover. Pronunciation assets and installed dependencies retain their respective terms; the Inflect license does not replace those terms or grant voice/identity rights.
For questions, use model discussions, GitHub or Discord. For a reproducible issue, include the revision, voice, device and a short text example. Keep credentials and private recordings out of public reports.
- Downloads last month
- 256