owensong/Inflect-Mini-v3-Preview

Inflect V3 Mini Preview

Experimental quiet research preview, not production-ready. No promotional launch is planned. This is not the finished Inflect V3 release and is not recommended for production, safety-critical tasks, or major deployments. Try it to inspect and report failures, not as a promise of stable quality, successor release dates, or future API compatibility. Atom is not included.

Six fixed English speakers: cassidy, hope, james, jasper, mark, sophia (speaker IDs zero through five in that order). Output is mono 24 kHz audio.

Inflect V3 Preview collection contains the Micro, Nano and Mini previews. This package is intended for owensong/Inflect-Mini-v3-Preview.

Join The Inflect Discord

Join the Inflect community on Discord to follow development, hear about future releases, share listening feedback, and discuss integrations. If you want to help shape what comes next, this is the place to join.

Found a clipped word, strange pitch, or a pronunciation failure? Share the exact text, model, voice, seed/settings and a short generated clip. Those concrete examples are more useful than an aggregate score. Please do not post API keys or private recordings.

Listen First

Six different examples cover conversational turns, narration and practical messages. The same texts are used across Micro, Nano and Mini so you can compare them. These are illustrative preview demos, not benchmark or human-MOS results. The delivery is generated by the unchanged preview: pitch/timing, word-boundary and synthetic-timbre problems can still be audible.

Each example is a single predetermined render at the native settings, not the best of a seed sweep. Complete audio is kept without cleanup, trimming, gain adjustment or repair. The labels describe the text, not trained emotion/style controls. Settings, all attempts and hashes are in demos/manifest.json.

Hope / Conversation: clarification

Wait, did you mean tomorrow morning? I thought we were meeting tonight. Let me check the message again.

Full WAV | 6.1 seconds

Cassidy / Short narration

The rain had finally stopped. She opened the window, took a deep breath, and listened to the city waking up.

Full WAV | 6.5 seconds

Mark / Application message

The update is ready to install. Save your work first, then restart the application when you have a moment.

Full WAV | 6.2 seconds

Listen to James, Jasper and Sophia

James / Conversational reassurance

I know that was frustrating. We can take this one step at a time, and I'll help you work through it.

Full WAV | 6.9 seconds

Jasper / Everyday instructions

The train leaves in fifteen minutes. If we take the next bus, we should still reach the station on time.

Full WAV | 6.1 seconds

Sophia / Longer sentence and pauses

Before you send it, check the name, the address, and the final paragraph. A small correction now could save some trouble later.

Full WAV | 7.9 seconds

Run

Use Python 3.11 or newer and install requirements.txt in your environment. Choose a suitable PyTorch CPU/CUDA wheel for your machine. This folder includes all neural weights and runtime sources: inference never downloads a model and does not import the original project checkout. The eSpeak shared library is provided by the separately installed espeakng-loader dependency.

Download the complete repository once with Hugging Face Hub, then run locally. No account is required to download this public preview.

python -m pip install "huggingface-hub>=0.23.4,<1"

from huggingface_hub import snapshot_download

snapshot_download(
    "owensong/Inflect-Mini-v3-Preview",
    local_dir="Inflect-Mini-v3-Preview",
)

Run the following commands from the downloaded Inflect-Mini-v3-Preview folder:

python -m pip install -r requirements.txt
python infer.py --text "The morning light shines through the window." --voice hope --output hope.wav
python smoke_test.py

CPU is the default. --device cuda selects an installed local CUDA device. Defaults match the development screen: 16 Euler flow steps, temperature 0.667, seed 1234. Seeds are repeatable within the tested CPU environment, not a promise of bit-identical audio across devices or software versions. No reference audio, alignment, voice cloning, event controls, or extra model is required. The CLI and smoke test write raw FLOAT WAVs without PCM16 quantization, gain, normalization, fades, denoising or clipping correction. Numeric settings must be finite; steps are integer 1..64 and seeds are integer 0..2**63-1.

Original six-voice packaging checks (same sentence)

Original Packaging Samples

Each sample says: "The morning light shines through the window." These are the existing CPU smoke outputs: raw mono 24 kHz FLOAT WAVs, 16 steps, temperature 0.667, seed 1234, with no post-processing or sample cherry-picking. They demonstrate functional inference, not stable voice traits, production quality, or a completed human-listening gate. Listen for the known preview defects rather than relying on an ASR or predicted-quality score.

Cassidy raw WAV

Hope raw WAV

James raw WAV

Jasper raw WAV

Mark raw WAV

Sophia raw WAV

Checkpoint And Size

Selected checkpoint: spectrotemporal-full-selected-blend65k, SHA-256 6502bccca4faf974b2cfd34420994b8899d3662218da8a9d903f7664c0301d28. This is a 75% full-data 60k / 25% full-data 80k weight interpolation, not a checkpoint trained for exactly 65k updates. It is evaluation-only and not optimizer-resumable. weights/acoustic.pt contains only its model tensors; exported tensors were checked individually for exact equality with the source. The BigVGAN base generator is separately pinned and unmodified.

Deployment Component	Exact Parameters
Text encoder and duration predictor	4,202,981
Speaker embeddings	384
Predicted prosody	585,219
Spectrotemporal flow estimator	8,835,972
Acoustic subtotal	13,624,556
BigVGAN base vocoder, weight norm removed	14,015,041
Complete deployed graph	27,639,597

These are instantiated graph counts, not rounded model labels or optimizer/ buffer tensor counts. The selected acoustic model has two additional scalar normalization buffers. The vocoder's stored weight-normalization representation has a different tensor count; the deployment total is after removing weight norm. config.json contains the exact model, mel, frontend, and speaker contracts. provenance.json preserves the blend's parent hashes, training-data receipts, source-code hashes, export equality result, and copied-asset receipts. Historical provenance paths are evidence, never inference dependencies.

Evidence, Not A Release Win

The last full overnight selection used 48 reused development prompts across six voices, not an untouched test set. Three UTMOSv2 fusion_stage3 scoring runs, each configured with three internal native repetitions (num_repetitions=3), averaged 3.398980 selected vs 3.366482 original (+0.032498). These are three scoring-run aggregates, not three single native predictions or nine independently reserved evaluation panels. UTMOSv2 is predicted quality, not human MOS. Averaged per-item deltas split 24 positive and 24 negative; the approximate 95% paired interval is [-0.021412, 0.086408]. It crosses zero. Selection on this panel makes the apparent gain optimistic.

On the same screen, faster-whisper base.en counted 23 vs 25 word errors out of 1,021 reference words (2.2527% vs 2.4486%). This is scoped ASR evidence, not independently qualified intelligibility or a comparison against other TTS systems. Both used predicted timing, 16 steps and temperature 0.667. The original reports and per-item scores are included in evidence/. The 48 original and 48 selected benchmark WAVs are bundled without any audio conversion under evidence/original-vs-selected/, with filenames aligned to the reports and a self-contained audio_manifest.json. Every WAV was resolved by its recorded SHA-256, not guessed from a filename. Historical source-report hashes remain evidence receipts; they are not runtime filesystem dependencies. No benchmark was rerun or relabeled during packaging.

UTMOS22 was not established for this selected blend on this panel and is not substituted from another Mini checkpoint. These scores must not be compared directly with Micro's fresh one-voice matched48 or Nano's different panels.

The local smoke/report.json is a CPU functional check of one short prompt per speaker with sockets blocked: finite, nonempty, non-clipped waveform, correct 24 kHz/hop contract, exact parameter ledgers, and invalid-input rejection. Its observed timings are diagnostic only; they are not a throughput, quality, latency, or real-time claim. See the report for tested dependency versions.

Preview Limitations

• Buzz, high-frequency artifacts, pronunciation and pacing defects can remain; six-voice human listening is still the release gate, not completed by this smoke test.
• English prose only; multilingual synthesis, expressive events, cloning, streaming and long-form stability are not qualified.
• Spell out numbers. Digits, currency, markup and audio-event syntax are rejected instead of being silently normalized by an unqualified frontend.
• Inputs are limited to 512 phoneme/blank positions and 2,000 predicted mel frames. Split longer prose yourself; automatic chunking is not included.
• The exact eSpeak 1.52.0 contract is checked. Other frontend versions may pronounce text differently and are rejected rather than treated as equivalent.
• This package has not been certified on all Python, OS or accelerator versions.

Licenses And Publication Gate

The owner designates Inflect-authored code and selected Mini acoustic preview weights Apache-2.0, supplied in LICENSE. Vendored Matcha and BigVGAN code retain their separate MIT/Apache attribution texts in licenses/. The official BigVGAN base model card identifies that generator's model license as MIT. The GPL frontend dependencies are not copied or relicensed as Apache; they remain separately installed. See THIRD_PARTY_NOTICES.md and LICENSE_AUDIT.md for the component/dependency inventory.

The owner reports contacting Fish, ElevenLabs and Inworld and having permission for training and release. Those permissions are owner-reported, not independently audited here. This preview license designation is not an independent legal clearance of provider terms or voice/data rights. Third-party licenses remain in effect. Packaging QA is technical qualification, not proof that human-listening or final production-quality gates passed. No upload, rental change, or other model modification was performed by packaging.

build_package.py --workspace PATH is a local source-audit/export helper only. It needs original assets to rebuild the package; neither inference nor smoke testing imports it. It does not perform downloads or publishing. PACKAGING_QA.json records technical preview qualification; manifest.json lists package file sizes and SHA-256 hashes. Verify with python finalize_qa.py --verify after downloading the complete folder.

Help Improve The Next Release

Join the Inflect Discord for development updates and to share reproducible failures, comparisons and integration feedback. This remains a research preview, not a production endorsement. Thank you for testing it honestly.