React Native ExecuTorch v0.11.1
What's New?
🗣️ Reverb TTS with voice cloning
- Reverb: multilingual text-to-speech (24 kHz) that speaks in the voice of any 1–20 s recording and its transcript, cloned on-device with
encodeVoice()(#1638). - 7 languages (en, de, es, fr, it, pl, pt) and 16 preset voices.
- Runs on MLX + Core ML (iOS), QNN or Vulkan (Android), with XNNPACK as the fallback.
useTextToSpeech(models.textToSpeech.REVERB.DEFAULT), orcreateReverbTextToSpeech.
⚡ Initial Qualcomm QNN support
- New opt-in
qnnbackend runs models on the Hexagon NPU of Snapdragon 8 Gen 1 and newer, Android arm64 only (#1482). - Add
"qnn"tobackendsin yourreact-native-executorchconfig;DEFAULTthen picks the QNN variant built for the device's Hexagon version, and falls back to Vulkan or XNNPACK elsewhere. - QNN variants: EfficientNetV2-S, DeepLabV3, FCN, LR-ASPP, SSDLite, YOLO26, RF-DETR Nano, Depth Anything V2/V3, ZipDepth, Whisper and Reverb. EfficientNetV2-S runs in 2.4 ms vs 46.9 ms on XNNPACK fp32 (Galaxy S26 Ultra).
- Apps need
useLegacyPackaging = true.
🐛 Fixes
- Native libs moved to
v0.11.1-libs, with Vulkan fixes on Adreno and Mali (#1637). - Options passed as
undefinedfall back to their defaults, including LLM sampling options (#1611, #1612). - SDXS frames the prompt with CLIP's start and end tokens (#1635).
- Gemma 4 Omni Vulkan re-exported to fix its memory use (#1628).
- iOS privacy manifest restored (#1627).
- The React Native 0.83 floor is declared as a peer dependency (#1625).
- Clear error naming the opencv lib when an app opted out of it (#1624).
What's Changed
- fix(llm): fall back to sampling defaults for options passed as undefined by @RaphaelFakhri in #1611
- fix: fall back to defaults for options passed as undefined by @msluszniak in #1612
- feat(core): add fp16 dtype representation by @IgorSwat in #1632
- fix(cv): frame the SDXS prompt with CLIP's start and end tokens by @Rasaboun in #1635
- fix(libs): bump native libs to v0.11.1-libs by @IgorSwat in #1637
- fix: declare the react-native 0.83 floor as a peer by @msluszniak in #1625
- fix(ios): restore the privacy manifest dropped in 0.10 by @msluszniak in #1627
- fix(cv): name the opencv lib when an app opted out of it by @msluszniak in #1624
- fix: use v1.0.0 Gemma Omni Vulkan model by @barhanc in #1628
- feat(android): add the Qualcomm QNN backend by @msluszniak in #1482
- feat(speech): add reverb-tts pipeline by @IgorSwat in #1638
Full Changelog: v0.11.0...v0.11.1