Inflect v3 Preview
Collection
Six English voices. Mini 34.31M, Micro 12.63M, Nano 4.47M. Raw audio, measured results and automatic access. Preview, not final. • 3 items • Updated • 1
Inflect Micro v3 Preview
Preview release, not the final v3. Download access is granted automatically after a Hugging Face access request. Final licensing and distribution review are pending; no final-release date is announced. Access and licensing notes.
A smaller working footprint. Six English voices in a 12.63M-parameter inference core, with lower measured CPU compute than Mini.
Text goes in. A complete waveform comes out. Voice selection, text-to-acoustic prediction and waveform generation live in one deployable model package; there is no inference-time teacher or external learned vocoder.
Six voices · 24 kHz mono · local inference · raw FP32 output
Listen · Choose a model · Evaluation · Run locally · Technical notes
Listen
Hope and James illustrate this checkpoint. All six voices are available in the runtime. These are existing, unfiltered outputs from the packaged candidate, selected by prompt order rather than quality score. No EQ, denoising, pitch correction or experimental adapters were added for this presentation.
I brought an extra bowl because your neighbor mentioned that someone might join us for soup.
Hope
James
These demonstration clips are public; model downloads use the repository's access gate. If an embedded player is unavailable, use its download link.
The v3 family
| v3 Preview | Core parameters | FP32 weights | Choose it for |
|---|---|---|---|
| Mini | 34.31M | 138.39 MB | Quality-first |
| Micro | 12.63M | 51.62 MB | Lower compute and footprint |
| Nano | 4.47M | 18.93 MB | Smallest; experimental audio quality |
Micro trades some acoustic detail for a smaller model and lower measured CPU compute. All three contain Hope, Sophia, Cassidy, James, Jasper and Mark in one checkpoint; changing voices does not require a second model download.
Core counts include the text, acoustic and waveform-generation networks. Small fixed voice/style assets and external G2P software are additional. Weights are decimal MB, not total installation size or peak memory.
Evaluation
The evidence below comes from existing internal evaluations, not a new launch benchmark. Audio is unfiltered. WER is measured independently with Whisper large-v3 and Qwen3-ASR-1.7B; lower is better. UTMOS and UTMOS v2 are automated quality predictions, not human listening scores; higher is better.
Ordinary English: 96 clips per model
48 prompts, two voice slots, identical text. Every tested system completed all 96 ordinary cases. The panel covers conversation, questions, narration, instructions, long passages and phonetic coverage.
| Model / tested runtime | Whisper WER ↓ | Qwen WER ↓ | UTMOS ↑ | UTMOS v2 ↑ |
|---|---|---|---|---|
| Inflect Mini / FP32 | 0.000% | 0.048% | 4.475 | 3.931 |
| Inflect Micro / FP32 | 0.096% | 0.000% | 4.468 | 3.670 |
| Kokoro-82M / FP32 | 0.000% | 0.143% | 4.448 | 3.852 |
| Phonon 31M / q8 | 0.239% | 0.096% | 4.439 | 3.722 |
| Phonon 40M / q8 | 0.382% | 0.239% | 4.437 | 3.747 |
| Piper / ONNX | 0.334% | 0.382% | 4.324 | 3.628 |
| Kitten / SDK | 1.815% | 0.191% | 4.157 | 2.084 |
The tested voices differ: Inflect Hope/James, Kokoro Heart/Michael, Phonon Zoey/Easton, Kitten Bella/Jasper and Piper Lessac/Ryan. These are specific voices and runtime configurations, not every voice offered by each system.
Harder text exposes a remaining gap. On 24 additional stress prompts across two voices, Mini and Micro each rendered 34/48 attempts and refused 14/48. All five competitors rendered 48/48. These refusals are failures, not correct speech, and are reported separately from ordinary-text WER.
Read the comparison boundaries before interpreting the scores
WER is word-weighted across complete clips. English-formatting normalization was repaired retrospectively and applied uniformly across models; raw scores and error counts are retained. A zero on this panel is not a zero general error rate. Different voices, synthesis settings and runtime precision remain confounds. Phonon uses q8 while Inflect and native Kokoro use FP32. Automated quality predictors can disagree with listeners. This is not an independent leaderboard or a sealed generalization test, and we do not claim universal superiority.
Full methodology, stress results and coverage · Unrounded results and evidence hashes
Existing CPU deployment measurements
| Model | RTF ↓ | Audio / compute ↑ | Median complete-utterance latency |
|---|---|---|---|
| Inflect Mini / FP32 | 0.4910 | 2.04x | 2,568 ms |
| Inflect Micro / FP32 | 0.1816 | 5.51x | 895 ms |
| Kokoro-82M / FP32 | 0.5229 | 1.91x | 2,742 ms |
AMD EPYC 7K62, two physical cores, two tensor threads, batch one; eight warmups and 16 measured calls. Historical October 21 deployment test; raw-text timing includes G2P and excludes model loading. These are full-utterance results, not streaming latency, mobile results or a memory benchmark. Phonon's different q8/Rust deployment is documented separately in Evaluation.
Run locally
After automatic access approval, use your own Hugging Face account to download the package. Create an isolated Python environment first.
python -m pip install "huggingface_hub>=0.34"
hf auth login
hf download owensong/Inflect-Micro-v3 --local-dir inflect-micro-v3
cd inflect-micro-v3
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
Change --voice to hope, sophia, cassidy, james, jasper or mark.
The published CLI uses CPU FP32 with two tensor threads. Setup downloads the
frontend assets; subsequent inference can run offline. The export was exercised
with Python 3.12 and PyTorch 2.8; this is not a certified cross-platform installer.
The CLI accepts up to 2,000 characters and splits long text at bounded word
boundaries, inserting 120 ms between segments. This is not true streaming.
Explicit phoneme overrides are available with --phonemes. See
Technical notes for the frontend, controls and package checks.
Scope and limitations
Included: frozen base weights, six voice choices, integrated waveform generation, standalone inference code, a text CLI and versioned evidence. No experimental polish, filters or tag adapters are applied to this release.
- Questions and expressive phrasing can sound flatter than intended.
- Ambiguous words, names, abbreviations, numbers and symbols can be misread or rejected. Custom phonemes offer an explicit override, not an automatic fix.
- Texture, timing and voice similarity can differ from the teacher; metrics do not replace listening.
- Emotion/event tags, whisper, laughter, voice cloning, multilingual synthesis, true streaming and validated INT8/FP8 deployment are not supported features.
- No production-readiness or support guarantee is made. The hosted playground is not yet connected to this release; the examples above are prerecorded model outputs.
Built on Kokoro distillation
Inflect v3 uses Kokoro-based distillation and adapted voice conditioning. Mini retains a Kokoro-derived native waveform generator behind a smaller learned text/acoustic path; Micro and Nano use trained compact waveform generators. These are not models trained wholly from scratch, and they do not load the full Kokoro teacher during inference.
Upstream attribution is retained in Third-party notices. Final model-weight licensing and voice/data distribution review are incomplete. Gated access is not itself commercial-use permission; see Access notes.
Exact candidate and weight hashes are in candidate.json. Release manifest covers the package files. Preview revisions may change; pin a commit when reproducing results.
Feedback
Report pronunciation, missing words, unnatural phrasing, voice inconsistency or deployment problems in this repository's Community tab. Include the model revision, voice, exact text and environment; add an audio clip if possible. Do not post sensitive text, credentials or private voice references.
Maintained by Owen Song. Inflect v3 Preview is a base-model evaluation release, not the final v3 launch.