q3146dq4/Supertonic-3-LiteRT-WI8-AFP32

Supertonic-3 LiteRT WI8-AFP32-Selective PTQ

Based on soniqo/Supertonic-3-LiteRT.

Quantization

• vector_estimator.tflite: Dynamic WI8_AFP32
• vocoder.tflite: selective Dynamic WI8_AFP32; two sensitive vocoder ops are kept FP32
• text_encoder.tflite: FP32
• duration_predictor.tflite: Dynamic WI8-AFP32

Performance

Measured on a Snapdragon 690 Android device using the LiteRT XNNPACK CPU backend.

Inference steps	Original soniqo/Supertonic-3-LiteRT	WI8-AFP32	Speedup
4-step	RTF 0.667	RTF 0.305	2.19× faster
8-step	RTF 1.205	RTF 0.513	2.35× faster

Lower RTF is better. An RTF below 1.0 means synthesis is faster than real-time.

Compared with the original Soniqo LiteRT model, WI8-AFP32 reduces end-to-end RTF by approximately 54% at 4 steps and 57% at 8 steps on Snapdragon 690.

This allows even relatively low-end mobile CPUs to run Supertonic-3 comfortably faster than real-time at both 4-step and 8-step settings.

Performance varies depending on device, thermal state, thread configuration, and LiteRT/XNNPACK version.

This repository contains LiteRT/TFLite inference models only.

Duration predictor precision

duration_predictor.tflite is Dynamic WI8-AFP32. text_encoder.tflite remains FP32. In the local 4/8/16-step A/B comparison, the overlapping waveform/spectral metrics matched the previous FP32-DP bundle; the maximum predicted-duration shift was 11.96 ms.