Inflect v3 Preview
Collection
Six voices. Mini 34.31M, Micro 12.63M, Nano 4.47M. Raw audio, verified native-text benchmarks and automatic access. Preview, not final. • 3 items • Updated • 1
Inflect Mini v3 Preview
Preview release: not the final v3 launch.
We're sharing this evaluation release directly, one-to-one with testers rather than running a public launch or advertising campaign. Features, packaging and documentation may change before the main release. No final-release date is announced.
This page is public and discoverable, not invite-only. Downloads require a Hugging Face account and an access request; approval is automatic. These are base-speech weights. Experimental emotion and event tags are not included. Access and licensing.
Join the Inflect Discord for preview feedback, development updates and release news. Project GitHub
34.31M core parameters / Six English voices / 24 kHz mono
Quality-focused English speech in a compact, standalone model. The family artwork is conceptual; Atom is not released.
Listen / Choose a size / Sound quality / Word accuracy / Run locally
Six voices
Start with Cassidy and Mark. Every voice is available in this checkpoint. Different passages show different uses; the same voice reads the same passage across Mini, Micro and Nano.
01 / CONVERSATION
CassidyCould you bring the smaller suitcase downstairs? WAV 3.4s · raw 24 kHz |
02 / NARRATION
MarkBeyond the orchard, a narrow footpath curved toward the ruined tower. WAV 5.1s · raw 24 kHz |
03 / EXPLANATION
HopeA shadow becomes longer when the light source moves closer to the horizon. WAV 4.5s · raw 24 kHz |
04 / DIALOGUE
SophiaWait! I left something important on the counter. WAV 3.0s · raw 24 kHz |
05 / INSTRUCTIONS
JamesPress the side button once, then wait until the indicator turns green. WAV 5.2s · raw 24 kHz |
06 / LONGER PASSAGE
JasperWhen the rain stopped, we carried the chairs outside and wiped the table dry. Everyone helped with dinner, and by the time the sun went down, the garden was warm and quiet again. WAV 12.4s · raw 24 kHz |
Raw standalone text-to-speech from the packaged weights. No denoising, EQ, pitch correction, gain matching or teacher acoustic inputs. Existing benchmark clips; passages assigned by category, not score. Audio provenance.
Help shape v3. Join the Inflect Discord and tell us what you hear. Include the model, voice and exact text so we can reproduce it.
Choose your size
| Preview | Parameters | FP32 weights | Two-core CPU RTF | Role |
|---|---|---|---|---|
| Mini | 34.31M | 138.39 MB | 0.534 | Quality-focused |
| Micro | 12.63M | 51.62 MB | 0.206 | Lower compute and footprint |
| Nano | 4.47M | 18.93 MB | 0.091 | Smallest; experimental quality |
All three checkpoints contain Hope, Sophia, Cassidy, James, Jasper and Mark. Voice selection does not require downloading another model. FP32 weight sizes use decimal MB; they are not total installation size or peak RAM. The CPU timings are measured on the same rental CPU, not mobile-device predictions.
Benchmarks
96 ordinary-English texts · two voice slots · 192 clips per system. All systems read the same text through their native frontend. The main panel uses Hope/James for Inflect; the six-voice diagnostic below uses 12 texts per voice. These are internal development measurements, not a human listening leaderboard.
Sound quality
UTMOS v2 predicts perceived quality, not word accuracy or listener preference. Higher is better. The dot plot uses a zoomed 2.0-4.1 axis; the scores are unchanged.
Exact quality scores and predictor caveats
| Model | UTMOS v2 |
|---|---|
| Inflect Mini | 3.978 |
| Inflect Micro | 3.725 |
| Inflect Nano (20k) | 3.454 |
| Kokoro 82M | 3.883 |
| Phonon 31M / Q8 | 3.782 |
| Kitten micro 0.8 | 2.099 |
| Piper | 3.630 |
Fold 0, three case-seeded repetitions. Different stock speaker identities across systems. Mini's paired difference versus Kokoro is +0.094, with a conditional 95% text-cluster interval of [+0.066, +0.123]; this is not human preference. Kitten's v2 score disagrees with its original UTMOS of 4.175, so do not treat that low prediction as a conclusive quality verdict. Full protocol.
Word accuracy
Whisper large-v3 and Qwen3-ASR-1.7B decode independently, without reference-text prompting. Lower WER is better. Every ordinary row has 2,700 reference words per recognizer; rates are word-weighted, not averages of clip percentages.
| Model | Whisper WER | Qwen WER | Two-ASR Mean |
|---|---|---|---|
| Inflect Mini | 0.22% | 0.11% | 0.17% |
| Inflect Micro | 0.15% | 0.07% | 0.11% |
| Inflect Nano (20k) | 0.96% | 0.52% | 0.74% |
| Kokoro 82M | 0.30% | 0.07% | 0.19% |
| Phonon 31M / Q8 | 0.15% | 0.11% | 0.13% |
| Kitten micro 0.8 | 1.37% | 0.11% | 0.74% |
| Piper | 0.30% | 0.41% | 0.35% |
Two-ASR Mean = (unrounded Whisper WER + unrounded Qwen WER) / 2.
This is a descriptive average, not a third ASR result, an ensemble transcript or
a more accurate estimate of human intelligibility. The individual scores remain
visible because recognizers disagree. Mini's mean, for example, is
(6/2700 + 3/2700) / 2 = 0.1667%, displayed as 0.17%.
Differences of a few recognized words do not establish a reliable model ranking.
Unrounded results and error counts / Frozen text panel / Evaluation protocol
Stress text, completion rates and the Mini frontend repair
32 additional prompts cover formatted numbers, dates, times, units, symbols, names and difficult pronunciation. Refusals count as failures, not correct speech. WER below covers successfully scored clips only; the cohorts are unequal.
Sound quality
| Model | Scored / planned | UTMOS v2 |
|---|---|---|
| Inflect Mini | 48 / 64 | 3.918 |
| Inflect Micro | 48 / 64 | 3.662 |
| Inflect Nano (20k) | 48 / 64 | 3.303 |
| Kokoro 82M | 64 / 64 | 3.818 |
| Phonon 31M / Q8 | 64 / 64 | 3.677 |
| Kitten micro 0.8 | 64 / 64 | 2.051 |
| Piper | 64 / 64 | 3.602 |
Word accuracy
| Model | Scored / planned | Whisper WER | Qwen WER |
|---|---|---|---|
| Inflect Mini | 48 / 64 | 4.27% | 0.70% |
| Inflect Micro | 48 / 64 | 3.42% | 1.40% |
| Inflect Nano (20k) | 48 / 64 | 3.97% | 2.54% |
| Kokoro 82M | 64 / 64 | 3.12% | 2.13% |
| Phonon 31M / Q8 | 64 / 64 | 5.02% | 4.25% |
| Kitten micro 0.8 | 64 / 64 | 7.85% | 5.67% |
| Piper | 64 / 64 | 6.52% | 6.55% |
The strict-frontend run refused the same eight texts in both slots for each Inflect size. Mini's later fallback update accepts all 132 original panel/timing texts plus 24 new regression texts. Previously supported phonemes are unchanged, and eight audio controls are PCM-identical. That targeted check does not replace the original stress results or certify every proper-name pronunciation. Mini frontend qualification. Micro and Nano have not received this runtime change.
All six voices: a smaller diagnostic panel
The same 12 ordinary texts, two per category, were measured for every voice in every Inflect size. This is a small diagnostic subset, not 96 prompts per voice.
Sound quality
| Voice | UTMOS v2 |
|---|---|
| Hope | 4.000 |
| Sophia | 3.879 |
| Cassidy | 4.038 |
| James | 3.957 |
| Jasper | 3.929 |
| Mark | 3.982 |
Word accuracy
| Voice | Whisper WER | Qwen WER |
|---|---|---|
| Hope | 0.00% | 0.00% |
| Sophia | 0.60% | 0.00% |
| Cassidy | 0.00% | 0.00% |
| James | 0.00% | 0.00% |
| Jasper | 0.00% | 0.00% |
| Mark | 0.00% | 0.00% |
Original UTMOS / secondary predictor
| Model | Original UTMOS |
|---|---|
| Inflect Mini | 4.465 |
| Inflect Micro | 4.461 |
| Inflect Nano (20k) | 4.148 |
| Kokoro 82M | 4.426 |
| Phonon 31M / Q8 | 4.443 |
| Kitten micro 0.8 | 4.175 |
| Piper | 4.345 |
Same ordinary outputs. Secondary predictor, not human listening.
CPU runtime
The matched short/medium/long workload uses two distinct physical cores of an AMD EPYC 7K62, two requested runtime threads, one warmup and three measurements per text/voice case. RTF is generation time divided by generated audio time; below 1 is faster than real time. These are text-to-complete-waveform timings.
| Model | Tested runtime | Two-core RTF | Four-core RTF |
|---|---|---|---|
| Inflect Mini | PyTorch FP32 | 0.534 | 0.315 |
| Inflect Micro | PyTorch FP32 | 0.206 | 0.136 |
| Inflect Nano (20k) | PyTorch FP32 | 0.091 | 0.069 |
| Kokoro 82M | PyTorch FP32 | 0.679 | 0.405 |
| Phonon 31M / Q8 | Native Q8 | 0.174 | Not tested |
| Kitten micro 0.8 | Native SDK / ONNX | 0.122 | Not tested |
| Piper | Native SDK / ONNX | 0.199 | Not tested |
The aggregate uses the same 18 successful measured calls per system; paragraphs are excluded for every model because the evaluated Inflect frontend refused them. Phonon Q8 and PyTorch FP32 are different runtime/precision stacks. Kitten/Piper include their SDK WAV write/read. Loading, network transfer and downloads are excluded. This is not streaming, time-to-first-audio, a mobile benchmark or an architecture-only speed ranking.
Run locally
Request access on this page, sign in with your own Hugging Face account, then use an isolated Python environment:
python -m pip install "huggingface_hub>=0.34"
hf auth login
hf download owensong/Inflect-Mini-v3-Preview --local-dir inflect-mini-v3-preview
cd inflect-mini-v3-preview
python -m pip install -r requirements.txt
python -m spacy download en_core_web_sm
python synthesize.py --text "What would you like to hear next?" --voice cassidy --output example.wav
Choose hope, sophia, cassidy, james, jasper or mark. The supplied CLI
uses CPU FP32 with two tensor threads. Setup downloads dependencies; subsequent
inference can run offline. The tested environment is Linux, Python 3.12 and PyTorch
2.8, not a certified cross-platform installer.
The default CLI uses offline Misaki plus the bundled eSpeak unknown-word fallback. --strict-g2p reproduces the earlier strict path. Custom full phonemes use --phonemes and bypass G2P.
The CLI accepts up to 2,000 characters and inserts 120 ms between bounded text
segments. It never silently truncates unsupported phonemes. Runtime details.
What this preview does not promise
- Names, ambiguous words, numbers and symbols can still be mispronounced or refused.
- Question intonation and expressive phrasing can sound flatter than intended.
- The base packages do not support emotion/event tags, whisper, laughter, arbitrary voice cloning, multilingual synthesis or true streaming.
- INT8/FP8 deployment is not a supported release feature. Weight size is not peak RAM.
- Nano has more audible quality trade-offs; a low word-error rate does not establish crisp, natural sound. Listen before choosing a size.
Provenance and access
Inflect v3 is distilled from a voice-adapted Kokoro teacher and uses Kokoro/StyleTTS2-derived components. Inference runs independently, without the teacher. Third-party notices retain upstream attribution.
The core parameter count includes the learned text, acoustic and waveform-generation components; no separate learned vocoder is required. Fixed voice assets, G2P assets and the Python runtime are additional dependencies. Not every component was trained from scratch.
Final model-weight licensing and voice/data distribution review are pending. Automatic gated access is not a commercial-use license. Do not infer an Apache weight license, redistribution clearance or a production-support commitment from the artwork, code notices or access approval. Read the access notes.
This is the public-facing preview distribution, not the development workspace. Weights are pinned in candidate.json; the release manifest covers the package. Pin a repository revision when reproducing results.
Feedback
Join the Inflect Discord for listening feedback and preview updates. The project GitHub is the broader Inflect repository, not a complete public v3 source release.
Report an issue or listening observation with the revision, voice, exact text and environment. Audio helps. Please do not post credentials, sensitive text or private voice references.
Owen Song / Inflect v3 Preview. A broader release follows validation and feedback; no final-release date is announced.