phoonnx mirror of KittenML/KittenTTS — a family of tiny, fully offline, CPU-only English TTS models. The upstream ONNX graphs are used unmodified; this repo adds phoonnx-canonical config.json files (vocab + inference defaults) and splits each voices.npz into one style .bin file per voice, matching phoonnx's StyleTTS2Adapter engine_params contract.

Upstream sources (Apache-2.0, verified on each model card):

nano-0.1/model.onnx, config.json, <voice>.bin (x8)
nano-0.2/model.onnx, config.json, <voice>.bin (x8)
mini-0.1/model.onnx, config.json, <voice>.bin (x8)

Voices (same 8 across all tiers): expr-voice-2-m, expr-voice-2-f, expr-voice-3-m, expr-voice-3-f, expr-voice-4-m, expr-voice-4-f, expr-voice-5-m, expr-voice-5-f.

input_ids (int64) + style (1, 256) + speed (1) -> waveform @ 24kHz, the same single-graph contract phoonnx's StyleTTS2Adapter already implements for StyleTTS2 and Kokoro — no new adapter code was needed. Tokenization is a per-character map over espeak-ng en-us IPA (with stress marks), reproduced verbatim from kittentts==0.1.3's inline vocab, not hand-copied.

WER (parakeet-tdt-0.6b-v2, 5 sentences, expr-voice-2-f):

No torch reference implementation is published upstream (KittenTTS ships ONNX only), so there is no torch<->onnxruntime parity check to run. Upstream ships exactly one precision per tier (no separate fp16/int8 variants), so no variant-consistency check applies either.

No 5000-sample trailing trim is applied: kittentts==0.1.3's own KittenTTS.generate() returns the raw graph output with no post-processing, and inspecting the tail of phoonnx-synthesized audio shows no noise burst (tail RMS in the same range as the rest of the clip).