From 4c2ca8c82723161932fde85344bba9c579a947de Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Mateusz=20S=C5=82uszniak?= <mateusz.sluszniak@swmansion.com>
Date: Mon, 31 Aug 2026 12:30:47 +0200
Subject: [PATCH] feat(speech): add Vulkan whisper-tiny variants

Registers the fp16 and int8 Vulkan exports of multilingual whisper-tiny and
lets the speech demo pick them on Android.

The Vulkan encode needs the ExecuTorch 1.4.1 runtime from v0.10.0-libs-1.4.1:
earlier Vulkan libs lack the gelu tanh clamp, the reduce-shader barrier fix and
the Adreno local-work-group fix.
---
 apps/speech/app/audio-file-transcription/index.tsx | 10 ++++++++++
 apps/speech/app/microphone-transcription/index.tsx | 10 ++++++++++
 .../scripts/download-libs.js                       |  4 ++--
 packages/react-native-executorch/src/models.ts     | 14 ++++++++++++++
 4 files changed, 36 insertions(+), 2 deletions(-)

diff --git a/apps/speech/app/audio-file-transcription/index.tsx b/apps/speech/app/audio-file-transcription/index.tsx
index 1b58375c40..f6935f02b1 100644
--- a/apps/speech/app/audio-file-transcription/index.tsx
+++ b/apps/speech/app/audio-file-transcription/index.tsx
@@ -73,6 +73,16 @@ const MODELS = [
     config: models.speechToText.WHISPER.TINY.MLX_BF16,
     disabled: Platform.OS !== 'ios',
   },
+  {
+    name: 'Tiny Multilingual (Vulkan fp16)',
+    config: models.speechToText.WHISPER.TINY.VULKAN_FP16,
+    disabled: Platform.OS !== 'android',
+  },
+  {
+    name: 'Tiny Multilingual (Vulkan int8)',
+    config: models.speechToText.WHISPER.TINY.VULKAN_INT8,
+    disabled: Platform.OS !== 'android',
+  },
   {
     name: 'Base Multilingual (CPU)',
     config: models.speechToText.WHISPER.BASE.XNNPACK_FP32,
diff --git a/apps/speech/app/microphone-transcription/index.tsx b/apps/speech/app/microphone-transcription/index.tsx
index 1355e5991a..70ee8606de 100644
--- a/apps/speech/app/microphone-transcription/index.tsx
+++ b/apps/speech/app/microphone-transcription/index.tsx
@@ -66,6 +66,16 @@ const MODELS = [
     config: models.speechToText.WHISPER.TINY.MLX_BF16,
     disabled: Platform.OS !== 'ios',
   },
+  {
+    name: 'Tiny Multilingual (Vulkan fp16)',
+    config: models.speechToText.WHISPER.TINY.VULKAN_FP16,
+    disabled: Platform.OS !== 'android',
+  },
+  {
+    name: 'Tiny Multilingual (Vulkan int8)',
+    config: models.speechToText.WHISPER.TINY.VULKAN_INT8,
+    disabled: Platform.OS !== 'android',
+  },
   {
     name: 'Base Multilingual (CPU)',
     config: models.speechToText.WHISPER.BASE.XNNPACK_FP32,
diff --git a/packages/react-native-executorch/scripts/download-libs.js b/packages/react-native-executorch/scripts/download-libs.js
index 459cdcfba9..718a1dc64b 100644
--- a/packages/react-native-executorch/scripts/download-libs.js
+++ b/packages/react-native-executorch/scripts/download-libs.js
@@ -124,8 +124,8 @@ const FEATURE_MAP = {
   multimodalLLM: { backends: ['xnnpack', 'mlx', 'vulkan'], libs: ['opencv'] },
   // Privacy filter classifiers ship xnnpack + an MLX iOS export.
   privacyFilter: { backends: ['xnnpack', 'mlx'], libs: [] },
-  // Whisper ships xnnpack + coreml.
-  speechToText: { backends: ['xnnpack', 'coreml'], libs: [] },
+  // Whisper ships xnnpack + coreml + a Vulkan Android export.
+  speechToText: { backends: ['xnnpack', 'coreml', 'vulkan'], libs: [] },
   // Kokoro ships xnnpack only.
   textToSpeech: { backends: ['xnnpack'], libs: ['phonemis'] },
   // FSMN VAD — xnnpack only.
diff --git a/packages/react-native-executorch/src/models.ts b/packages/react-native-executorch/src/models.ts
index 1985444133..69784a8c97 100644
--- a/packages/react-native-executorch/src/models.ts
+++ b/packages/react-native-executorch/src/models.ts
@@ -750,6 +750,18 @@ const WHISPER_TINY_MLX_INT8: WhisperSttModel = {
   supportedLanguages: WHISPER_LANGUAGES,
   vadModel: FSMN_VAD_XNNPACK_FP32,
 };
+const WHISPER_TINY_VULKAN_FP16: WhisperSttModel = {
+  modelPath: `${BASE_URL}-whisper-tiny/${NEXT_VERSION_TAG}/vulkan/whisper_tiny_vulkan_fp16.pte`,
+  tokenizerPath: `${BASE_URL}-whisper-tiny/${NEXT_VERSION_TAG}/tokenizer.json`,
+  supportedLanguages: WHISPER_LANGUAGES,
+  vadModel: FSMN_VAD_XNNPACK_FP32,
+};
+const WHISPER_TINY_VULKAN_INT8: WhisperSttModel = {
+  modelPath: `${BASE_URL}-whisper-tiny/${NEXT_VERSION_TAG}/vulkan/whisper_tiny_vulkan_int8.pte`,
+  tokenizerPath: `${BASE_URL}-whisper-tiny/${NEXT_VERSION_TAG}/tokenizer.json`,
+  supportedLanguages: WHISPER_LANGUAGES,
+  vadModel: FSMN_VAD_XNNPACK_FP32,
+};
 
 const WHISPER_BASE_EN_XNNPACK_FP32: WhisperSttModel<'en'> = {
   modelPath: `${BASE_URL}-whisper-base.en/${NEXT_VERSION_TAG}/xnnpack/whisper_base_en_xnnpack_fp32.pte`,
@@ -1862,6 +1874,8 @@ export const models = {
         COREML_FP16: WHISPER_TINY_COREML_FP16,
         MLX_BF16: WHISPER_TINY_MLX_BF16,
         MLX_INT8: WHISPER_TINY_MLX_INT8,
+        VULKAN_FP16: WHISPER_TINY_VULKAN_FP16,
+        VULKAN_INT8: WHISPER_TINY_VULKAN_INT8,
       },
       /**
        * Multilingual Whisper Base model. Higher accuracy across supported