Inflect v3 Preview
Collection
Six English voices. Three local TTS sizes: 34.31M, 12.63M, 4.47M. Raw samples and verified benchmarks. Automatic gated preview; not final. • 3 items • Updated • 1
Public preview, not the final v3 release. Automatic gated downloads; base speech only, without experimental tags. Access terms and pending licensing review.
Inflect Micro v3 Preview
Every voice. A smaller footprint.
English text to 24 kHz speech. Local inference. No inference-time teacher.
12,632,966 core parameters · 51.62 MB FP32 weights · 6 voices
Listen / Benchmarks / Run locally / Choose a size / Collection
Micro balances audio quality with lower storage and measured CPU compute. Its text, acoustic and compact waveform generator are packaged together.
Hope, Sophia, Cassidy, James, Jasper and Mark are together in one checkpoint. Text prediction, acoustic prediction and waveform generation are included. No separate learned vocoder or inference-time teacher is required.
What the parameter and download sizes include
Core counts include the text, acoustic and waveform-generation networks. Fixed voice/style assets and external G2P software are additional. FP32 weight sizes are decimal MB, not installation size or peak RAM. See Technical notes for the exact package boundaries.
Listen
Micro examples show Hope and James; all six voices are included. These are unchanged outputs from the packaged checkpoint, selected by prompt order rather than quality score. No filters, enhancement, polish or tag adapters were added. Audio examples are public even before download access.
I brought an extra bowl because your neighbor mentioned that someone might join us for soup.
Sample provenance and hashes. If your reader cannot display the players, use the WAV links. These are recorded outputs, not a connected live playground.
The v3 family
| Model | Core parameters | FP32 weights | Role |
|---|---|---|---|
| Mini Preview | 34.31M | 138.39 MB | Quality-first |
| Micro Preview | 12.63M | 51.62 MB | Compact balance |
| Nano Preview | 4.47M | 18.93 MB | Smallest; experimental quality |
The figure uses the same 96-clip ordinary panel below. It is an internal learned quality estimate, not human MOS. Nano is the packaged 20k checkpoint; its later experimental continuations and filters are not part of this preview.
Evaluation
64 texts. Seven systems. Two recognizers. UTMOS v2 estimates audio quality; Whisper large-v3 and Qwen3-ASR-1.7B check the words separately. Ordinary speech and difficult text remain separate evaluations. No human preference is claimed.
Frontend boundary: this research comparison uses shared Misaki plus an opt-in eSpeak fallback for five unsupported text families. That fallback is not the shipped CLI default, and Kokoro here is its core with the shared frontend, not stock KPipeline. See coverage and protocol.
Ordinary English / 96 clips per system
Same 48 texts, two configured voices per system. Parentheses are errors/reference words; WER is word-weighted. All rows below have 96 scored outputs in this derived fallback-enabled evaluation, not a packaged-CLI completion guarantee.
| Model / tested runtime | UTMOS v2 ↑ | Whisper WER ↓ | Qwen WER ↓ |
|---|---|---|---|
| Inflect Mini / FP32 | 4.007 | 0.58% (5/866) | 0.81% (7/866) |
| Inflect Micro / FP32 | 3.788 | 0.69% (6/866) | 0.69% (6/866) |
| Inflect Nano / FP32, 20k | 3.512 | 3.12% (27/866) | 1.15% (10/866) |
| Kokoro core / FP32 | 3.847 | 0.81% (7/866) | 0.35% (3/866) |
| Phonon 31 / q8 | 3.854 | 0.58% (5/866) | 0.46% (4/866) |
| Kitten micro 0.8 / SDK | 2.092 | 1.96% (17/866) | 0.69% (6/866) |
| Piper / ONNX | 3.666 | 0.81% (7/866) | 0.46% (4/866) |
Specific voices, not each competitor's entire voice catalog: Inflect Hope/James; Kokoro Heart/Michael; Phonon Zoey/Easton; Kitten Bella/Jasper; Piper Lessac/Ryan. Speaker identities differ. Phonon uses q8; the others use the stated FP32/SDK/ONNX stack. Piper metrics use a labeled 24 kHz proxy of native 22.05 kHz audio.
The unusually low Kitten UTMOS v2 result disagrees with its original UTMOS. Do not interpret it as a matching human-preference gap. This is not an independently administered leaderboard or a sealed generalization test.
Two-ASR mean / optional descriptive summary
For readers who want one summary, this is the arithmetic mean of the two
unrounded, word-weighted normalized WERs: (Whisper WER + Qwen WER) / 2.
Ordinary and stress panels are never pooled.
| Model | Ordinary two-ASR mean | Stress two-ASR mean |
|---|---|---|
| Inflect Mini / FP32 | 0.69% | 7.71% |
| Inflect Micro / FP32 | 0.69% | 8.81% |
| Inflect Nano / FP32, 20k | 2.14% | 9.82% |
| Kokoro core / FP32 | 0.58% | 7.70% |
| Phonon 31 / q8 | 0.52% | 10.94% |
| Kitten micro 0.8 / SDK | 1.33% | 9.21% |
| Piper / ONNX | 0.64% | 8.98% |
A mean is not a more accurate ground-truth WER. Recognizers can share errors, mishear correct speech, or accept incorrect speech. This does not correct either transcript, estimate human intelligibility, or establish independent agreement. Both original recognizer scores and their counts remain visible above. We do not choose whichever recognizer gives a model the lowest error rate.
Difficult text / 32 matched clips per system
Same 16 stress texts, two voices. Includes structured numbers and difficult verbalizations. Normalized WER uses predeclared accepted spoken references; reference lengths can differ. Raw-reference counts remain in the evidence.
| Model / tested runtime | UTMOS v2 ↑ | Whisper WER ↓ | Qwen WER ↓ |
|---|---|---|---|
| Inflect Mini / FP32 | 3.882 | 8.55% (47/550) | 6.88% (38/552) |
| Inflect Micro / FP32 | 3.582 | 9.09% (50/550) | 8.53% (47/551) |
| Inflect Nano / FP32, 20k | 3.334 | 10.00% (55/550) | 9.64% (53/550) |
| Kokoro core / FP32 | 3.728 | 8.73% (48/550) | 6.68% (37/554) |
| Phonon 31 / q8 | 3.658 | 12.41% (68/548) | 9.46% (53/560) |
| Kitten micro 0.8 / SDK | 2.005 | 9.54% (54/566) | 8.88% (49/552) |
| Piper / ONNX | 3.656 | 10.95% (60/548) | 7.02% (40/570) |
The original strict 832 ledger had 780 successes, 42 refusals and 10 failures. The derived 832 overlays 52 actual frontend rescues; those failures are not erased. External Kitten/Piper add 256 retained outputs, giving 1,088 total output records. That total is not 1,088 newly rendered clips.
Micro across all six voices / same 16 ordinary texts
Sixteen clips/voice, 96 total. This overlaps the primary panel; do not pool it into a larger independent test. No competitor-wide six-voice claim is implied.
| Voice | Clips | UTMOS v2 ↑ | Whisper WER ↓ | Qwen WER ↓ |
|---|---|---|---|---|
| Hope | 16 | 3.712 | 1.37% (2/146) | 1.37% (2/146) |
| Sophia | 16 | 3.873 | 2.05% (3/146) | 1.37% (2/146) |
| Cassidy | 16 | 3.689 | 1.37% (2/146) | 2.74% (4/146) |
| James | 16 | 3.840 | 1.37% (2/146) | 2.74% (4/146) |
| Jasper | 16 | 3.786 | 2.05% (3/146) | 2.74% (4/146) |
| Mark | 16 | 3.848 | 1.37% (2/146) | 2.74% (4/146) |
Original UTMOS / secondary quality predictor
Retained for transparency, not mixed with UTMOS v2 or treated as human MOS.
Ordinary 96 per system:
| Model / tested runtime | Original UTMOS ↑ |
|---|---|
| Inflect Mini / FP32 | 4.443 |
| Inflect Micro / FP32 | 4.451 |
| Inflect Nano / FP32, 20k | 4.112 |
| Kokoro core / FP32 | 4.408 |
| Phonon 31 / q8 | 4.430 |
| Kitten micro 0.8 / SDK | 4.156 |
| Piper / ONNX | 4.363 |
Stress 32 per system:
| Model / tested runtime | Original UTMOS ↑ |
|---|---|
| Inflect Mini / FP32 | 4.497 |
| Inflect Micro / FP32 | 4.483 |
| Inflect Nano / FP32, 20k | 4.237 |
| Kokoro core / FP32 | 4.464 |
| Phonon 31 / q8 | 4.445 |
| Kitten micro 0.8 / SDK | 4.066 |
| Piper / ONNX | 4.298 |
Uncertainty, normalization and earlier comparisons
Paired 95% bootstrap intervals use 10,000 fixed-seed text-cluster samples, keeping voice pairs together. Mini-minus-Kokoro ordinary UTMOS v2: +0.159 [0.110, 0.206]; Whisper WER difference: -0.231 percentage points [-0.825, +0.233]. These describe this panel, not universal superiority. No human preference percentage is claimed for v3.
Both raw and normalized errors, complete hypotheses, chosen references and source hashes are retained. Phonon 40 appears only in the earlier October 22 evaluation; it was not run in this newer panel and is not spliced into its table.
Full evaluation protocol · Unrounded summary · Per-output evidence
CPU deployment / matched FP32
| Model / FP32 | Whole-utterance RTF ↓ | Audio / compute ↑ | Median latency |
|---|---|---|---|
| Inflect Mini | 0.494 | 2.02x | 1,534 ms |
| Inflect Micro | 0.196 | 5.10x | 567 ms |
| Inflect Nano | 0.083 | 12.03x | 249 ms |
| Kokoro core | 0.494 | 2.02x | 1,594 ms |
Same server, two physical cores, two Torch threads, batch 1. Four short prompts; three warmups and eight measured calls per prompt. Warm G2P is included; model loading is excluded. These are complete-utterance results, not streaming latency, laptop performance, a quantized speed comparison or a RAM benchmark.
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.