Inflect Kokoro Voices

A collection of fine-tuned Inflect-Micro-v2 voice models. Each voice lives in its own self-contained top-level folder (~73 MB) containing both a PyTorch checkpoint (model.pth) and ONNX exports (onnx/duration.onnx + onnx/decode.onnx), so a single folder serves both runtimes. All voices synthesize 24 kHz mono audio.

Voices

Voices below have been retrained with the v2 recipe (boundary-trimmed training data + Nyquist antialias filter in inference.py). Snippets render live from current weights. Remaining legacy voices are being retrained incrementally.

Quick start

Download a single voice (folders are independent — no shared files at the repo root are required at runtime):

from huggingface_hub import snapshot_download

snapshot_download(
    "Danny-Dasilva/inflect-kokoro-voices",
    allow_patterns="sky/*",
    local_dir=".",
)

Then synthesize:

python sky/inference.py --model-dir sky --text "Hello from the sky voice." --output out.wav

Repository layout

inflect-kokoro-voices/
├── README.md
├── LICENSE
├── THIRD_PARTY_NOTICES.md
├── samples/                    # one sample clip per voice (24 kHz wav)
└── sky/                        # one folder per voice, fully self-contained
    ├── model.pth               # PyTorch checkpoint
    ├── onnx/
    │   ├── duration.onnx       # phoneme -> durations
    │   └── decode.onnx         # durations + text -> waveform
    ├── inference.py            # reference inference script (PyTorch + ONNX)
    ├── config.json
    ├── symbols.json
    ├── frontend.json
    ├── runtime/                # minimal model + text-frontend code
    └── requirements.txt

Model details

• Architecture: VITS-family Inflect-Micro-v2, 9.36M parameters
• Audio: 24 kHz, mono
• Per-voice full model: each voice is a complete fine-tuned model, not a style embedding — voice switching means loading a different folder
• Runtimes: PyTorch (model.pth) and ONNX (onnx/), exported from the same checkpoint

Training

Each voice is fine-tuned on ~5.1 hours of synthetic audio generated from a 3600-sentence phonetically-balanced corpus, for 28,000 steps with the balanced preset, on a single RTX 5090.

Evaluation

The sky voice's training configuration (5.1h data / 28k steps) was selected via a 50-sentence out-of-domain paired evaluation against two ablations (2.26h/14k and 5.1h/14k):

The released configuration wins speaker similarity significantly (paired Wilcoxon p < 1e-4 vs both ablations, bootstrap 95% CI excludes 0), with parity on naturalness/intelligibility metrics; 0/50 catastrophic clips (WER > 0.5) in all systems. Fleet voices use the same recipe.

Training data & provenance

The training audio for these voices is synthetic, generated with hexgrad/Kokoro-82M (Apache-2.0) as a data-generation teacher. No Kokoro weights are included in, or loaded by, these models — Kokoro was used only to render the training corpus; the model weights here descend solely from Inflect-Micro-v2.

Kokoro's own training data includes permissively licensed corpora whose attribution requirements are carried forward here:

• Koniwa (tnc voice data) — CC-BY-3.0
• SIWIS — CC-BY-4.0

See THIRD_PARTY_NOTICES.md for full attributions.

Limitations

• Synthetic-voice provenance: these voices are distillations of synthetic voices; they do not clone any real, identifiable person. The source voices derive from Kokoro's openly licensed synthetic voicepacks.
• English only: trained exclusively on American English text and audio; other languages are unsupported.
• Prosody and expressiveness are bounded by the synthetic teacher (teacher ceiling); highly emotive or long-form dramatic reading may sound flat.

License & attribution

Weights and code in this repository are released under Apache-2.0 (see LICENSE). Third-party components and carried-forward data attributions (VITS, BigVGAN, alias-free-torch, Kokoro-82M, Koniwa CC-BY-3.0, SIWIS CC-BY-4.0) are documented in THIRD_PARTY_NOTICES.md.