Inflect v3 Preview
Collection
Six voices each. Mini 34.31M, Micro 12.63M, Nano 4.47M, DSP-based Atom 157K synthesis parameters. Automatic access. Preview, not final. • 4 items • Updated • 1
Access the Inflect v3 Preview
Non-commercial evaluation preview. Approval is automatic. By accessing the covered materials, you agree to the Inflect Preview Evaluation License 1.0; no redistribution or hosted services. Upstream and earlier grants remain intact.
Log in or Sign Up to review the conditions and access this model content.
Inflect Atom v3 Preview
The smallest voice in the Inflect family.
Six English voices. 24 kHz speech. Run it locally.
PREVIEW RELEASE
This is not the final Inflect v3 launch.
Shared directly with testers ahead of a wider release, without a public launch campaign. These pages are public and discoverable. Download approval is automatic after sign-in.
Non-commercial evaluation only. No redistribution or hosted services.
Preview license.
Final-release terms are undecided. Earlier valid grants and upstream licenses remain intact.
157,190 synthesis-core parameters · 0.66 MB FP32 checkpoint
Six fixed voices · 24 kHz mono · Offline synthesis after setup
Six voices
Start with Cassidy and Mark. Then hear the rest of the family. Each voice has its own short passage; that passage stays the same across all four models.
CONVERSATION
Cassidy
Could you bring the smaller suitcase downstairs?
INSTRUCTIONS
Mark
Press the side button once, then wait until the indicator turns green.
EXPLANATION
Hope
A shadow becomes longer when the light source moves closer to the horizon.
NARRATION
James
Beyond the orchard, a narrow footpath curved toward the ruined tower.
DIALOGUE
Sophia
Wait! I left something important on the counter.
DIFFICULT-WORD STRESS TEST
Jasper
Photosynthesis, crystallography, and electromagnetic interference demand precise measurements.
Original 24 kHz samples, without denoising, EQ, or gain matching. Audio downloads and checksums are available in the sample manifest.
A voice for your device
Choose the footprint that fits your application, then listen before deciding. All four previews offer the same six fixed voice IDs, local synthesis after setup, and a complete waveform output. No reference recording is needed.
| Preview | Synthesis-core parameters | FP32 checkpoint | Profile |
|---|---|---|---|
| Atom | 157,190 | 0.66 MB | Ultra-small core |
| Nano | 4,465,702 | 18.93 MB | Neural synthesis |
| Micro | 12,632,966 | 51.62 MB | Neural synthesis |
| Mini | 34,310,790 | 138.39 MB | Neural synthesis |
All four include six voices in one checkpoint. Synthesis-core counts exclude frontend and fixed voice assets; checkpoint size is not installation size or peak RAM. Complete model accounting.
Atom uses a separate pronunciation frontend: the reference configuration has 1,727,704 active learned parameters including the 157,190-parameter synthesis core. The 0.66 MB figure is the core checkpoint, not the complete installation.
Building with Inflect? Join the Discord. Share a short sample, tell us what you are making, or report a sentence that did not sound right. Tester feedback is shaping the final release.
Listen first. Then look closer.
Speech quality, word accuracy, and speed answer different questions. We report them separately and retain the counts and protocols behind the summaries. These results describe the published preview checkpoints, not unreleased fast-model experiments.
DSP matched-text panel / 2026-10-24. 48 shared text families; Atom six voices / 288 clips, five rivals one voice / 48 each. 528/528 core requests succeeded. Mini, Micro and Nano are unmeasured on this corpus.
Sound quality / UTMOS v2
Higher is better. This is predicted naturalness, not a human listening score. The dot plot uses a clearly labeled 2.0-4.5 truncated axis; values outside it are marked.
Exact scores, coverage, and full-range view
| System | UTMOS v2 | Core scored / planned | 95% family interval |
|---|---|---|---|
| Inflect Atom / six voices | 2.810 | 288/288 | 2.775-2.844 |
| Vocetta2-147K | 1.045 | 48/48 | 1.009-1.080 |
| Vocetta2-276K | 1.304 | 48/48 | 1.268-1.340 |
| Vocetta2-500K | 2.236 | 48/48 | 2.161-2.311 |
| Sano Heart-Nano | 1.388 | 48/48 | 1.343-1.428 |
| AuroraTTS-100K | 3.096 | 48/48 | 3.040-3.154 |
See the 0-5 view and uncertainty notes.
Word accuracy / averaged WER
Lower is better. The average uses separately computed, word-weighted corpus WER on the same audio within each model: two recognizers for this panel. It is a summary of ASR errors, not a human-verified speech error rate. Small score differences should not be read as proof that one model is more intelligible.
| System | ASR-mean WER | Paired core clips |
|---|---|---|
| Inflect Atom / six voices | 1.53% | 288 |
| Vocetta2-147K | 3.08% | 48 |
| Vocetta2-276K | 3.22% | 48 |
| Vocetta2-500K | 2.65% | 48 |
| Sano Heart-Nano | 2.58% | 48 |
| AuroraTTS-100K | 3.49% | 48 |
Recognizer results, word counts, categories, and disagreements
WER = (substitutions + deletions + insertions) / reference words. The displayed mean is calculated from unrounded corpus rates, not per-sentence percentages. Shared normalization handles formatting, not missing word endings or changed meaning. Recognizer disagreements can reflect the ASR, the reference, or the speech itself.
Individual engines and counts / Category and stress results / Evaluation protocol.
Local performance / CPU
Native text to complete audio, including pronunciation conversion, on two physical AMD EPYC 7K62 cores. Twelve texts, one excluded warmup, three measured repetitions per text and voice; weights are loaded before timing.
| Published Inflect model | Median RTF | Audio seconds / wall second |
|---|---|---|
| Inflect Atom | 0.0481 | 20.8x |
| Inflect Nano | 0.0805 | 12.4x |
| Inflect Micro | 0.1896 | 5.3x |
| Inflect Mini | 0.5159 | 1.9x |
Audio seconds per wall second is inverse median RTF, not speedup over a competitor. Atom uses the CPU-LUT backend. Results exclude model loading and file writes; complete-utterance timing is not first-audio latency. Tested voices differ by system.
Competitor comparisons and complete benchmark records
Full CPU comparison and precision / Additional native-text models / Complete evaluation record.
The expanded comparison includes Atom, Nano, Micro, Mini, Kokoro, Phonon, Kitten, Piper, Supra, Silencio, Steer, and Paradee where qualified. Separate corpora and recognizer sets are not merged into a single ranking.
GPU performance / original Mini, Micro, and Nano
An additional RTX 5090 pass measures the original published FP32 checkpoints, including pronunciation conversion and complete waveform transfer. Twelve texts, two voices, 72 measured trials per model; CUDA synchronized at both boundaries.
| Original preview | Median RTF | Audio seconds / wall second | P50 seconds | Trials |
|---|---|---|---|---|
| Inflect Mini | 0.0349 | 28.7x | 0.151 | 72 |
| Inflect Micro | 0.0422 | 23.7x | 0.183 | 72 |
| Inflect Nano | 0.0388 | 25.8x | 0.165 | 72 |
Not a speedup over another model or first-audio latency. Atom is not measured on GPU. Protocol and all trial records.
Run locally
Read the preview license, request access, and use your own Hugging Face account. The download includes the current license and unchanged preview weights.
Request access, sign in with your own account, and use an isolated Python 3.12 environment.
python -m pip install "huggingface_hub>=0.34"
hf auth login
hf download owensong/Inflect-Atom-v3-Preview --local-dir inflect-atom-v3-preview
cd inflect-atom-v3-preview
python -m pip install -r requirements-cpu.txt
python -m spacy download en_core_web_sm
python synthesize.py --voice cassidy --text "What would you like to hear next?" --backend cpu-lut --output example.wav
Voice IDs: cassidy, mark, hope, james, sophia, jasper. Install the expected
en_core_web_sm 3.8.0 assets before offline use. See deployment notes for tested environments and frontend requirements.
Before you build
This preview provides fixed-voice English speech. Voice cloning, emotion/event tags, and true streaming are not included. Names, numbers, ambiguous words, and some inputs can be mispronounced or refused. Listen to outputs before using them. No production-readiness or cross-platform compatibility guarantee is made.
Do not impersonate people, misrepresent consent, or use the preview commercially. Disclose synthetic speech where it could mislead. Use non-sensitive test text. Supported scope and deployment details.
Preview terms and credits
Inflect Preview Evaluation License 1.0 applies to the identified Inflect materials: non-commercial evaluation, no model redistribution, and no hosted services. This is a preview-only license, not the final-release license or an open-source license. Full terms / Component scope / Access notes.
Upstream components retain their own licenses. Kokoro and StyleTTS2 are credited in third-party notices; no endorsement is implied. Earlier valid grants are preserved. Model, voice, and data rights are distinct; access approval does not itself clear third-party rights.
Version identity and technical records
Weight SHA-256: aad761146483f24f2669223e54bc66db6cfb869da9a4f2fdc969b7e03f9f8e96.
Model details / Evidence /
Benchmark files. No model weights, voices, or runtime code were
changed by this presentation and licensing update.
Help shape Inflect v3.
Join the Discord ·
Report an issue ·
Explore all four previews