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
Inflect Mini v3 Preview
Six English voices. One standalone checkpoint.
The quality-focused member of the Inflect family.
Unannounced preview, not the final v3 launch. Shared one-to-one with testers; these pages are public and discoverable. Sign in to Hugging Face and request access; download approval is automatic. Base speech only; emotion/event tags are not included. Licensing review is pending. Current preview terms.
34,310,790 synthesis-core parameters · 138.39 MB FP32 checkpoint
Six fixed voices · 24 kHz mono · Offline synthesis after setup
UTMOS v2 3.978 · ASR-mean WER 0.27%
Automatic metrics on the evaluation panel below.
Listen / Run locally / Model details / Evaluation.
Six voices
Start with Cassidy and Mark, then compare all six. Each voice reads a different sample, with the same text for that voice across all four sizes.
| Voice | Exact published input | Listen |
|---|---|---|
| Cassidy Conversation |
Could you bring the smaller suitcase downstairs? | WAV |
| Mark Instructions |
Press the side button once, then wait until the indicator turns green. | WAV |
| Hope Explanation |
A shadow becomes longer when the light source moves closer to the horizon. | WAV |
| James Narration |
Beyond the orchard, a narrow footpath curved toward the ruined tower. | WAV |
| Sophia Dialogue |
Wait! I left something important on the counter. | WAV |
| Jasper Difficult-word stress test |
Photosynthesis, crystallography, and electromagnetic interference demand precise measurements. | WAV |
Raw published samples, without denoising, EQ or gain matching. Audio provenance and checksums.
Benchmarks
Native-text panel / 2026-10-23. 96 ordinary texts x two voice slots = 192 clips per system. Hope/James for Inflect; stock identities differ. Atom was not included in this original seven-system run; its later native96 result is shown separately below.
Predicted naturalness
UTMOS v2 predicts naturalness; it is not human MOS. Higher is better. The dot plot uses a truncated 2.0-4.5 axis; arrows identify any off-scale scores.
| System | UTMOS v2 | Core scored / planned |
|---|---|---|
| Inflect Mini | 3.978 | 192/192 |
| Inflect Micro | 3.725 | 192/192 |
| Inflect Nano (20k) | 3.454 | 192/192 |
| Kokoro 82M | 3.883 | 192/192 |
| Phonon 31M / Q8 | 3.782 | 192/192 |
| Kitten micro 0.8 | 2.099 | 192/192 |
| Piper | 3.630 | 192/192 |
Full-range UTMOS v2 plot / 0-5 axis
Same scores and systems as the zoomed dot plot. Off-scale values are not omitted.
ASR-mean WER
Lower is better. Mean corpus WER from four separately run recognizers, Whisper, Qwen, Granite and Parakeet, on identical clips; a descriptive summary, not human accuracy. Adding judges does not guarantee greater accuracy. Individual engines and protocol are below.
| System | ASR-mean WER | Paired core clips |
|---|---|---|
| Inflect Mini | 0.27% | 192 |
| Inflect Micro | 0.19% | 192 |
| Inflect Nano (20k) | 0.93% | 192 |
| Kokoro 82M | 0.31% | 192 |
| Phonon 31M / Q8 | 0.13% | 192 |
| Kitten micro 0.8 | 0.56% | 192 |
| Piper | 0.44% | 192 |
Advanced: individual ASR engines, counts, protocol and hashes
96 ordinary texts x two voice slots = 192 clips per system. Hope/James for Inflect; stock identities differ. Atom was not included in this original seven-system run; its later native96 result is shown separately below.
UTMOS v2 fold 0, three case-seeded repetitions. Separately run Whisper large-v3 and Qwen3-ASR-1.7B, no reference-text prompting. Word-weighted normalized WER; accepted references frozen before synthesis. Granite Speech 5.0 2B and NVIDIA Parakeet TDT 0.6B v3 add two separately run recognizers; the four-engine mean uses identical paired clips and the same frozen normalizer, without reference-text hints. More judges do not guarantee greater accuracy.
| System | Whisper WER | Qwen WER | Granite Speech 5.0 2B WER | Parakeet TDT 0.6B v3 WER | ASR-mean WER |
|---|---|---|---|---|---|
| Inflect Mini | 0.22% (6/2700) | 0.11% (3/2700) | 0.67% (18/2700) | 0.07% (2/2700) | 0.27% |
| Inflect Micro | 0.15% (4/2700) | 0.07% (2/2700) | 0.41% (11/2700) | 0.11% (3/2700) | 0.19% |
| Inflect Nano (20k) | 0.96% (26/2700) | 0.52% (14/2700) | 1.00% (27/2700) | 1.22% (33/2700) | 0.93% |
| Kokoro 82M | 0.30% (8/2700) | 0.07% (2/2700) | 0.67% (18/2700) | 0.22% (6/2700) | 0.31% |
| Phonon 31M / Q8 | 0.15% (4/2700) | 0.11% (3/2700) | 0.15% (4/2700) | 0.11% (3/2700) | 0.13% |
| Kitten micro 0.8 | 1.37% (37/2700) | 0.11% (3/2700) | 0.30% (8/2700) | 0.48% (13/2700) | 0.56% |
| Piper | 0.30% (8/2700) | 0.41% (11/2700) | 0.37% (10/2700) | 0.67% (18/2700) | 0.44% |
Inflect Mini: Whisper 0.22% (6/2700); Qwen 0.11% (3/2700); Granite Speech 5.0 2B 0.67% (18/2700); Parakeet TDT 0.6B v3 0.07% (2/2700). All four recognizers cover exactly the same 192 clip identities. Corpus WER is total substitutions + deletions
- insertions divided by total reference words, not a mean of per-clip or category rates.
ASR-mean WER is the arithmetic mean of the four unrounded corpus rates, not pooled
counts, an ensemble transcript, true accuracy, or human intelligibility.
Paired clip-ID set SHA-256:
b530daaff1f32381f37b3ae0a6561c1cecc3523c8770d9e2a191c19f9290be36(sorted IDs, same canonical JSON encoding described in evidence).
Normalizer: extended_tts_english_formatting_20261022_v3. Case/diacritic/apostrophe folding, enumerated unambiguous
contractions, grouped thousands and a.m./p.m. formatting; numeric values preserved.
Spoken alternatives were frozen before synthesis; no post-hoc synonyms or model-specific
exceptions. ASR-selected alternatives can yield different reference-word denominators.
Corpus SHA-256: 787ff8b779c20f45c315355b321a512651c1da04843bb4b84b287f440fe2040e.
Canonical JSON of all 128 frozen prompts: json.dumps(indent=2, ensure_ascii=True, allow_nan=False) plus LF; not an original file hash.
Metric-source SHA-256: 5d605a1faf7c7a9579cc06c5f0a214823ae0823c9793cbfe96a194d49aeba3f5.
| Additional recognizer | Pinned revision | Decoding |
|---|---|---|
| ibm-granite/granite-speech-5.0-2b | c075cf19e041a2e4d16c5ee24763e5b037d8f479 |
BF16, generic audio prompt, greedy / 512 new-token cap |
| nvidia/parakeet-tdt-0.6b-v3 | 541d1f99c6b0c3cd0b11a95167540bb8edefd82b |
FP32 native greedy TDT |
Both use 16 kHz mono input, no gain adjustment or VAD. Four-ASR cohort/category counts, six pairwise disagreement audits and source/codec receipts. Normalized transcript audit SHA-256: 7d10e6dd52d65ecd4a0d418be37c190c0019c7f855fa4cb0d2de07eb7eb6906d. Full transcripts are retained with evaluator outputs, not duplicated in this card overlay. Historical Whisper/Qwen-only diagnostics remain separately labeled.
| Original recognizer | Executed revision | Decoding |
|---|---|---|
| Whisper large-v3 | edaa852ec7e145841d8ffdb056a99866b5f0a478 |
CUDA FP16, English transcription, beam 5, temperature 0; no VAD, history or initial prompt |
| Qwen3-ASR-1.7B | 7278e1e70fe206f11671096ffdd38061171dd6e5 |
CUDA BF16, English, empty context, deterministic, batch 1, SDPA / 256 new-token cap |
Original per-batch inference receipts were reconciled against the frozen clip IDs, audio hashes and hypotheses. This verifies the executed revision and decoding contracts, not the hashes of model files loaded during those historical runs: those per-run asset hashes were not recorded. Both original recognizers used two CPU threads. Receipt-archive SHA-256: 6795fa5d881b700f316e54ee425def4ff7683a509b3ea1f9e2e3a0a581a5c23c.
Full counts, canonical/raw scoring boundaries, sample-rate handling and limitations: evaluation, benchmarks, machine-readable evidence.
Full benchmarks, category counts and stress failures / Evaluation protocol.
CPU speed
Native text to complete waveform, including pronunciation conversion. Twelve fixed ordinary texts, one excluded warmup and three measured trials per text/voice, on two physical AMD EPYC 7K62 CPU cores. Models are loaded before timing. Native voices differ; this is not a matched-speaker comparison.
| 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 |
The inverse of median RTF expresses generated audio seconds per wall second; it is not a speedup over another model. Full competitor table, precision and timing protocol.
RTF = wall time / generated audio duration; lower is better. The median uses all measured trials, not a pooled duration ratio. P50/P95 describe complete utterances, not first-audio latency or streaming. Loading, warmups, safety probes and artifact writes are excluded. CPU clocks were not fixed; this was a shared rental host, not exclusive hardware. The earlier 18-call table and private Fast Micro phoneme-only timings use different protocols and are not merged here.
Expanded comparison: more models, the same 96 texts
Paradee / ordinary native96
The original seven native-text systems and Paradee 1.1's stock af_heart FP32/INT8 profiles use the same 96 ordinary text families. Native voices and clip multiplicities differ; this is not a speaker-identity match.
| System / profile | Native voices / clips | UTMOS v2 |
|---|---|---|
| Inflect Mini | 2 / 192 | 3.978 |
| Inflect Micro | 2 / 192 | 3.725 |
| Inflect Nano (20k) | 2 / 192 | 3.454 |
| Kokoro 82M | 2 / 192 | 3.883 |
| Phonon 31M / Q8 | 2 / 192 | 3.782 |
| Kitten micro 0.8 | 2 / 192 | 2.099 |
| Piper | 2 / 192 | 3.630 |
| Inflect Atom | 2 / 192 | 2.733 |
| Supra 1.0 Preview | 1 / 96 | 3.091 |
| Silencio | 2 / 192 | 3.052 |
| Steer / flow + Vocos | 1 / 96 | 3.375 |
| Paradee 1.1 / FP32 | 1 / 96 | 3.512 |
| Paradee 1.1 / INT8 | 1 / 96 | 3.501 |
| System / profile | ASR-mean WER (4 engines) |
|---|---|
| Inflect Mini | 0.27% |
| Inflect Micro | 0.19% |
| Inflect Nano (20k) | 0.93% |
| Kokoro 82M | 0.31% |
| Phonon 31M / Q8 | 0.13% |
| Kitten micro 0.8 | 0.56% |
| Piper | 0.44% |
| Inflect Atom | 0.18% |
| Supra 1.0 Preview | 0.20% |
| Silencio | 0.51% |
| Steer / flow + Vocos | 0.13% |
| Paradee 1.1 / FP32 | 0.24% |
| Paradee 1.1 / INT8 | 0.28% |
Paradee methods, coverage and limitations
UTMOS v2 is a learned predictor, not human MOS. ASR-mean WER is the unrounded arithmetic mean of four separately computed, word-weighted corpus WERs on identical audio within each profile. Whisper, Qwen, Granite and Parakeet are separately run recognizers; four judges do not guarantee greater accuracy or independent errors. Frozen accepted references and normalization are unchanged. Inspect disagreements such as suit case / suitcase and spoken numbers rather than treating WER as ground truth.
Paradee uses freshly generated Misaki/eSpeak phones and its released stock ONNX graphs. FP32/INT8 share text/phoneme policy but have separate ONNX internal noise; no cross-precision PCM or GPU-noise-seed equivalence is claimed. No parameter count or speed is inferred. Each profile retains all 128 renders; only 96 ordinary clips are scored. Its 32 stress cases are rendered, unscored, with no invented stress metric. Old 28-case cached-phone Paradee measurements and historical Atom48 remain separate.
Atom enters this table only through an optional, four-judge qualified native96 rival bundle; Atom48 never supplies a row. Optional rivals with only two judges remain in their separately labeled table. Original native Whisper/Qwen execution-contract receipts are required to match revisions and decoding; historical model-file hash limitations remain documented in the original panel methods.
Plan SHA-256: 904b23b50eef33cab573ef2d678f47dbe3cf10746f73746dbe3021070bc95df1. Sanitized counts and source pins. Detailed evidence.
Choose your size
| Preview | Synthesis-core parameters | FP32 checkpoint | Design |
|---|---|---|---|
| Atom | 157,190 | 0.66 MB | DSP / experimental |
| Nano | 4,465,702 | 18.93 MB | Conventional neural |
| Micro | 12,632,966 | 51.62 MB | Conventional neural |
| Mini | 34,310,790 | 138.39 MB | Conventional neural |
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.
Run locally
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-Mini-v3-Preview --revision 4c90a59f746336a2befdfc83af13f9f626ae8b65 --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 --voice cassidy --text "What would you like to hear next?" --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.
Limitations and responsible use
Names, numbers and ambiguous words can be mispronounced or refused; intonation can be flat. Listen before choosing a size. English fixed voices only, without arbitrary voice cloning or true streaming. Supported scope and limitations.
Disclose synthetic speech where it could mislead. Do not impersonate real people, misrepresent consent, or submit sensitive text to an untrusted deployment. Inspect generated audio before relying on it.
Provenance and preview terms
Distilled from a voice-adapted Kokoro teacher; standalone inference does not require that teacher. Learned text, acoustic and waveform components are counted in the core; Kokoro/StyleTTS2-derived components retain upstream attribution. The existing repository terms and notices remain unchanged. Proposed new-release terms are a draft, not adopted for these previews. Access approval is not commercial-use or redistribution clearance. Current access terms / current notices / model and data details.
Frozen reference revision: 4c90a59f746336a2befdfc83af13f9f626ae8b65.
Weight SHA-256: 8460e7d54a2fbf7fb54b3c7be7569c493c852de9a2aa1a0fd9fd9a099de581ed.
Candidate. Join Discord or report an issue
with revision, voice, exact text and environment; omit credentials and private references.