# Inflect v3 - a owensong Collection — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/556059bb41285a01

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-17 04:29 ingested 2026-11-18 02:42

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xyb3a7/inflect_v3_a_owensong_collection/)

## retained text



```
Inflect v3: compact local text-to-speech. Six English voices, three model sizes, reference-voice tuning, event tags and optional 48 kHz output. Apache 2.0.

Inflect is a family of compact English text-to-speech models. Choose a voice, enter your text, and generate a WAV locally. Start with **Mini** for the most features, **Micro** for a smaller neural model, or **Atom** for the smallest learned synthesis core.

* **GitHub** : [https://github.com/owenawsong/Inflect](https://github.com/owenawsong/Inflect)
* **Space** : [https://huggingface.co/spaces/owensong/Inflect-v3](https://huggingface.co/spaces/owensong/Inflect-v3)

# Choose your model

[](https://github.com/owenawsong/Inflect#choose-your-model)

||[**Mini**](https://huggingface.co/owensong/Inflect-Mini-v3)|[**Micro**](https://huggingface.co/owensong/Inflect-Micro-v3)|[**Atom**](https://huggingface.co/owensong/Inflect-Atom-v3)|
|:-|:-|:-|:-|
|Best starting point|**Quality and features**|Compact neural speech|Smallest learned core|
|Learned synthesis core|**34.31M**|**11.79M**|**157.19K**|
|English preset voices|6|6|6|
|Native output|24 kHz|24 kHz|24 kHz|
|Optional 48 kHz enhancement|Yes|Not included|Not included|
|Reference-voice tuning|Yes|Not included|Not included|
|Optional vocal events|8 tags, all 6 voices|Not included|Not included|
|Download|[Mini package](https://huggingface.co/owensong/Inflect-Mini-v3/tree/main)|[Micro package](https://huggingface.co/owensong/Inflect-Micro-v3/tree/main)|[Atom package](https://huggingface.co/owensong/Inflect-Atom-v3/tree/main)|

These are **learned synthesis-core counts**, not total installation size or RAM. Pronunciation assets, dependencies, voice conditioning and optional components are additional. Atom uses signal-processing-based waveform generation rather than a separate learned neural vocoder.

# Benchmarks

[](https://github.com/owenawsong/Inflect#benchmarks)

Measured speech quality, intelligibility and speed answer different questions. The following results describe **published Mini native 24 kHz base speech**, not tuned voices, event performances or enhanced output.

|Measure|Mini result|What was measured|
|:-|:-|:-|
|**Average WER**|**0.739%**|1,688 Open TTS evaluation texts per voice, all six voices; self-run|
|**UTMOSv2**|**3.917**|180 clips, 30 per voice|
|**UTMOS22Strong**|**4.479**|The same 180-clip panel|
|**Warm GPU throughput**|**162.1x real time**|RTX 5090, optimized CUDA; small 12-render probe|
|**Warm CPU throughput**|**1.81x real time**|Ryzen 9 3900X, 2 Torch threads; small 12-render probe|

The WER result is **not an official leaderboard placement or an independent held-out evaluation**. Quality predictors are not human ratings. CPU and GPU results use different systems and runtime paths; they are not a controlled hardware speedup comparison or a latency guarantee.
```



document fetched 2026-11-18 00:46 · canonical https://reddit.com/r/LocalLLaMA/comments/1xyb3a7/inflect_v3_a_owensong_collection
