Modelpublic

soniqo/Supertonic-3-LiteRT

sourceHugging Faceopenrailupdated 23d agoView on Hugging Face

3likes447downloads

Model Card

SupertonicTTS-3 — LiteRT (.tflite, Android / Qualcomm NPU)

First-party LiteRT export of Supertonic-3's four non-autoregressive flow-matching graphs. Built by our own pipeline (speech-models/stmodels): weights lifted from the `Supertone/supertonic-3` ONNX initializers → PyTorch nn.Module → `litert_torch.convert` (torch.export → StableHLO → TFLite). This avoids the onnx2tf NCHW/ConvNeXt layout failures that block direct ONNX→TFLite for this model.

Graphs & parity (FP32, vs ONNX Runtime)

Module	tflite	parity max\
`duration_predictor.tflite`	3.4 MB	4.1e-05 ✓
`vector_estimator.tflite` (ODE denoiser)	244 MB	5.6e-03 ✓
`vocoder.tflite`	97 MB	2.6e-04 ✓
`text_encoder.tflite`	34 MB	1.1e-01 (localized; mean ~2.5e-4) ⚠️
`vector_estimator_L128.tflite` (L=128 bucket)	244 MB	1.7e-02 (relative 2.7e-3) ✓
`vocoder_L128.tflite` (L=128 bucket)	97 MB	2.7e-04 ✓

Shapes are fixed (T=128 text tokens; latent length L per graph) — litert_torch cannot lower the relpos attention + ConvNeXt stack with a symbolic L, so the latent axis ships in buckets: the base pair is L=64 (≈4.5 s of audio), and *_L128.tflite is the same two L-dependent graphs exported at L=128 (≈9 s). Pick the smallest bucket whose window holds a piece's predicted duration — short sentences run on the cheap L=64 pair, a long sentence is generated in one pass on L=128 instead of being split (speech-core's LiteRTSupertonicTts does this automatically and loads the L=128 pair on first use). The host runs the flow-matching ODE loop (vector_estimator ×total_steps). Assets to drive them: tts.json, unicode_indexer.json (G2P-free tokenizer), voice_styles/*.json.

Graph inputs are bound by tensor name (serving_default_args_N, N = signature position): the converter permutes input slots differently per toolchain version, so do not rely on slot order.

Running on Android / Qualcomm NPU

—CPU/GPU: LiteRT (ai_edge_litert / TFLite) interpreter with XNNPACK/GPU delegate.
—Qualcomm HTP/NPU: the LiteRT QNN delegate at runtime, or compile to a QNN context binary via Qualcomm AI Hub (qai_hub) from these graphs (static shapes are HTP-friendly). int8/int4 PTQ via ai-edge-quantizer for full HTP residency is a follow-up.

Attribution & license

—Weights: derivative of `Supertone/supertonic-3` (commit 3cadd1ee6394adea1bd021217a0e650ede09a323), Supertone Inc., arXiv:2503.23108 — OpenRAIL-M (use-based restrictions carry over).

Other Supertonic-3 formats

—**Supertonic-3 — ONNX (INT8)** — server / desktop (ONNX Runtime).
—**Supertonic-3 — CoreML** — iOS / Apple Neural Engine (.mlpackage).

Ecosystem

—**soniqo.audio** — website / use-case explorer (transcription, voice cloning, live ASR, voice agents).
—**speech-core** — C++ orchestration library; Supertonic plugs in as a TTSInterface LiteRT model.
—**speech-swift** — Apple Silicon MLX + CoreML runtime.
—**speech-android** — Android SDK consuming on-device LiteRT bundles.

Other LiteRT models in this collection

—VoxCPM2 — LiteRT (INT8) (TTS)
—Parakeet TDT 0.6B v3 — LiteRT (INT8) · Nemotron Speech Streaming — LiteRT (ASR)
—Silero VAD v5 — LiteRT (VAD)