software-mansion/react-native-executorch-supertonic

supertonic

This repository hosts the supertonic models exported for the React Native ExecuTorch library as ExecuTorch .pte programs, ready to run on device.

Variants

Path	Component	Backend	Precision
mlx/duration_predictor_mlx_fp32.pte	duration_predictor	mlx	fp32
mlx/text_encoder_mlx_fp32.pte	text_encoder	mlx	fp32
mlx/vector_estimator_mlx_fp32.pte	vector_estimator	mlx	fp32
mlx/vocoder_mlx_fp32.pte	vocoder	mlx	fp32
vulkan/duration_predictor_vulkan_fp16.pte	duration_predictor	vulkan	fp16
vulkan/text_encoder_vulkan_fp16.pte	text_encoder	vulkan	fp16
vulkan/vector_estimator_vulkan_fp16.pte	vector_estimator	vulkan	fp16
vulkan/vocoder_vulkan_fp16.pte	vocoder	vulkan	fp16
xnnpack/duration_predictor_xnnpack_fp32.pte	duration_predictor	xnnpack	fp32
xnnpack/text_encoder_xnnpack_fp32.pte	text_encoder	xnnpack	fp32
xnnpack/vector_estimator_xnnpack_fp32.pte	vector_estimator	xnnpack	fp32
xnnpack/vector_estimator_xnnpack_int8.pte	vector_estimator	xnnpack	int8
xnnpack/vocoder_xnnpack_fp32.pte	vocoder	xnnpack	fp32
xnnpack/vocoder_xnnpack_int8.pte	vocoder	xnnpack	int8

XNNPACK int8

xnnpack/vector_estimator_xnnpack_int8.pte and xnnpack/vocoder_xnnpack_int8.pte are 8da8w (int8 dynamic activations, int8 per-channel weights) on the Linear layers, with the ConvNeXt pointwise convs moved onto Linears. The int8 set reuses the fp32 duration predictor and text encoder. LibriSpeech test-clean texts, voice F1, 8 steps, WER of a whisper base.en round trip:

	WER (57 short)	WER (8 long)
fp32	3.36	3.32
fp32, other noise seed	3.26	3.80
int8	3.74	3.01

Unshipped backends

A backend directory carrying a NOTES.md and no .pte was exported and not published. Its note records what went wrong.

• coreml (why)

Repository structure

config.json                                  32 B
coreml/NOTES.md                              879 B
mlx/config.json                              6.7 kB
mlx/duration_predictor_mlx_fp32.pte          3.4 MB
mlx/text_encoder_mlx_fp32.pte                34.4 MB
mlx/vector_estimator_mlx_fp32.pte            244 MB
mlx/vocoder_mlx_fp32.pte                     96.7 MB
unicode_indexer.json                         271 kB
voice_styles/F1.json                         285 kB
voice_styles/F2.json                         286 kB
voice_styles/F3.json                         284 kB
voice_styles/F4.json                         285 kB
voice_styles/F5.json                         285 kB
voice_styles/M1.json                         285 kB
voice_styles/M2.json                         285 kB
voice_styles/M3.json                         283 kB
voice_styles/M4.json                         285 kB
voice_styles/M5.json                         285 kB
voices/F1.bin                                50.5 kB
voices/F2.bin                                50.5 kB
voices/F3.bin                                50.5 kB
voices/F4.bin                                50.5 kB
voices/F5.bin                                50.5 kB
voices/M1.bin                                50.5 kB
voices/M2.bin                                50.5 kB
voices/M3.bin                                50.5 kB
voices/M4.bin                                50.5 kB
voices/M5.bin                                50.5 kB
vulkan/config.json                           4.1 kB
vulkan/duration_predictor_vulkan_fp16.pte    1.7 MB
vulkan/text_encoder_vulkan_fp16.pte          17.3 MB
vulkan/vector_estimator_vulkan_fp16.pte      122 MB
vulkan/vocoder_vulkan_fp16.pte               48.4 MB
xnnpack/config.json                          13.1 kB
xnnpack/duration_predictor_xnnpack_fp32.pte  3.4 MB
xnnpack/text_encoder_xnnpack_fp32.pte        34.5 MB
xnnpack/vector_estimator_xnnpack_fp32.pte    245 MB
xnnpack/vector_estimator_xnnpack_int8.pte    62.9 MB
xnnpack/vocoder_xnnpack_fp32.pte             96.7 MB
xnnpack/vocoder_xnnpack_int8.pte             36.8 MB

Compatibility

These files are published for the ExecuTorch v1.4.1 runtime. ExecuTorch gives no forward compatibility guarantee, so an older runtime may fail to load them.

To use them in React Native ExecuTorch, pass the model constant shipped in the library's model registry to the corresponding task pipeline. See the documentation.

To load these files in your own ExecuTorch runtime, read the compatibility note first.