msluszniak · Oct 21, 2026 · Oct 21, 2026 · Oct 21, 2026 · Oct 21, 2026 · Oct 22, 2026
diff --git a/apps/speech/app/audio-file-transcription/index.tsx b/apps/speech/app/audio-file-transcription/index.tsx
@@ -8,7 +8,12 @@ import {
   TouchableOpacity,
   TextInput,
 } from 'react-native';
-import { useSpeechToText, models, WHISPER_SAMPLE_RATE_HZ } from 'react-native-executorch';
+import {
+  useSpeechToText,
+  models,
+  QWEN3_ASR_LANGUAGES,
+  WHISPER_SAMPLE_RATE_HZ,
+} from 'react-native-executorch';
 import { decodeAudioData } from 'react-native-audio-api';
 
 import ScreenWrapper from '../../components/ScreenWrapper';
@@ -126,8 +131,24 @@ const MODELS = [
     config: models.speechToText.WHISPER.SMALL.MLX_INT8,
     disabled: Platform.OS !== 'ios',
   },
+  {
+    name: 'Qwen3-ASR 0.6B (CPU)',
+    config: models.speechToText.QWEN3_ASR.SIZE_0_6B.XNNPACK_8DA4W,
+  },
+  {
+    name: 'Qwen3-ASR 0.6B (MLX + Core ML)',
+    config: models.speechToText.QWEN3_ASR.SIZE_0_6B.MLX_COREML_INT4,
+    disabled: Platform.OS !== 'ios',
+  },
 ];
 
+// Qwen3-ASR detects the language itself; 'auto' leaves it to the model.
+const AUTO = 'auto';
+const languagesOf = (config: (typeof MODELS)[number]['config']): readonly string[] =>
+  'supportedLanguages' in config
+    ? config.supportedLanguages
+    : [AUTO, ...Object.keys(QWEN3_ASR_LANGUAGES)];
+
 function STTAudioContent() {
   const [selectedModel, setSelectedModel] = useState(MODELS[0]!);
   const [selectedLanguage, setSelectedLanguage] = useState<string>('en');
@@ -145,7 +166,7 @@ function STTAudioContent() {
     error: modelError,
   } = useSpeechToText(selectedModel.config);
 
-  const supportedLanguages = selectedModel.config.supportedLanguages;
+  const supportedLanguages = languagesOf(selectedModel.config);
 
   useEffect(() => {
     return () => {
@@ -167,7 +188,8 @@ function STTAudioContent() {
       setStatus('Transcribing...');
 
       let currentText = '';
-      const text = await transcribe(samples, { language: selectedLanguage as any }, (token) => {
+      const language = selectedLanguage === AUTO ? undefined : selectedLanguage;
+      const text = await transcribe(samples, { language: language as any }, (token) => {
         currentText += token;
         setAudioText(currentText);
       });
@@ -196,7 +218,7 @@ function STTAudioContent() {
     <ScrollView style={styles.container} contentContainerStyle={styles.content}>
       <View style={styles.card}>
         <ModelPicker
-          label="Whisper Model"
+          label="Speech-to-Text Model"
           options={MODELS.map((m) => ({
             label: m.name,
             value: m,
@@ -205,14 +227,14 @@ function STTAudioContent() {
           selectedValue={selectedModel}
           onValueChange={(m) => {
             setSelectedModel(m);
-            setSelectedLanguage(m.config.supportedLanguages[0]!);
+            setSelectedLanguage(languagesOf(m.config)[0]!);
           }}
         />
         <ModelStatus
           isReady={isSttReady}
           downloadProgress={downloadProgress}
           error={modelError ? modelError.message : null}
-          modelTypeLabel="Whisper model"
+          modelTypeLabel="Speech-to-text model"
         />
         {supportedLanguages.length > 1 && (
           <ModelPicker

diff --git a/apps/speech/app/microphone-transcription/index.tsx b/apps/speech/app/microphone-transcription/index.tsx
@@ -1,6 +1,11 @@
 import React, { useEffect, useRef, useState } from 'react';
 import { Platform, View, Text, StyleSheet, ScrollView, TouchableOpacity } from 'react-native';
-import { useSpeechToText, models, WHISPER_SAMPLE_RATE_HZ } from 'react-native-executorch';
+import {
+  useSpeechToText,
+  models,
+  QWEN3_ASR_LANGUAGES,
+  WHISPER_SAMPLE_RATE_HZ,
+} from 'react-native-executorch';
 import { AudioManager, AudioRecorder } from 'react-native-audio-api';
 import DeviceInfo from 'react-native-device-info';
 
@@ -119,8 +124,24 @@ const MODELS = [
     config: models.speechToText.WHISPER.SMALL.MLX_INT8,
     disabled: Platform.OS !== 'ios',
   },
+  {
+    name: 'Qwen3-ASR 0.6B (CPU)',
+    config: models.speechToText.QWEN3_ASR.SIZE_0_6B.XNNPACK_8DA4W,
+  },
+  {
+    name: 'Qwen3-ASR 0.6B (MLX + Core ML)',
+    config: models.speechToText.QWEN3_ASR.SIZE_0_6B.MLX_COREML_INT4,
+    disabled: Platform.OS !== 'ios',
+  },
 ];
 
+// Qwen3-ASR detects the language itself; 'auto' leaves it to the model.
+const AUTO = 'auto';
+const languagesOf = (config: (typeof MODELS)[number]['config']): readonly string[] =>
+  'supportedLanguages' in config
+    ? config.supportedLanguages
+    : [AUTO, ...Object.keys(QWEN3_ASR_LANGUAGES)];
+
 const isSimulator = DeviceInfo.isEmulatorSync();
 
 function STTContent() {
@@ -141,7 +162,7 @@ function STTContent() {
     error: modelError,
   } = useSpeechToText(selectedModel.config);
 
-  const supportedLanguages = selectedModel.config.supportedLanguages;
+  const supportedLanguages = languagesOf(selectedModel.config);
 
   const recorderRef = useRef<AudioRecorder | null>(null);
 
@@ -168,7 +189,8 @@ function STTContent() {
 
     (async () => {
       try {
-        const textStream = stream({ language: selectedLanguage as any });
+        const language = selectedLanguage === AUTO ? undefined : selectedLanguage;
+        const textStream = stream({ language: language as any });
         for await (const result of textStream) {
           setCommittedText(result.committed);
           setNonCommittedText(result.nonCommitted);
@@ -216,7 +238,7 @@ function STTContent() {
     <ScrollView style={styles.container} contentContainerStyle={styles.content}>
       <View style={styles.card}>
         <ModelPicker
-          label="Whisper Model"
+          label="Speech-to-Text Model"
           options={MODELS.map((m) => ({
             label: m.name,
             value: m,
@@ -225,14 +247,14 @@ function STTContent() {
           selectedValue={selectedModel}
           onValueChange={(m) => {
             setSelectedModel(m);
-            setSelectedLanguage(m.config.supportedLanguages[0]!);
+            setSelectedLanguage(languagesOf(m.config)[0]!);
           }}
         />
         <ModelStatus
           isReady={isSttReady}
           downloadProgress={downloadProgress}
           error={modelError ? modelError.message : null}
-          modelTypeLabel="Whisper model"
+          modelTypeLabel="Speech-to-text model"
         />
         {supportedLanguages.length > 1 && (
           <ModelPicker

diff --git a/docs/docs/02-extensions/speech/03-speech-to-text.md b/docs/docs/02-extensions/speech/03-speech-to-text.md
@@ -1,14 +1,15 @@
 ---
 title: Speech-to-Text (STT)
 slug: /extensions/speech-to-text
-description: 'Transcribe spoken audio in real time on-device in React Native using OpenAI Whisper with integrated Voice Activity Detection.'
+description: 'Transcribe spoken audio in real time on-device in React Native using OpenAI Whisper or Qwen3-ASR with integrated Voice Activity Detection.'
 keywords:
   [
     react native,
     speech to text,
     stt,
     asr,
     whisper,
+    qwen3-asr,
     voice recognition,
     transcription,
     vad,
@@ -20,7 +21,7 @@ keywords:
 
 # Speech-to-Text (STT)
 
-The Speech-to-Text extension transcribes spoken audio into text directly on-device using OpenAI's Whisper model paired with an integrated FSMN Voice Activity Detector (VAD).
+The Speech-to-Text extension transcribes spoken audio into text directly on-device using OpenAI's Whisper or Alibaba's [Qwen3-ASR](#qwen3-asr) paired with an integrated FSMN Voice Activity Detector (VAD).
 
 The pipeline supports two primary workflows:
 
@@ -202,22 +203,35 @@ const transcript = stt.transcribeWorklet(audioData, { language: 'en' });
 
 See [Worklets & Threading](../../03-core-and-advanced/06-worklets-and-threading.md) for details on worklet execution contexts and zero-copy host objects.
 
+## Qwen3-ASR
+
+[Qwen3-ASR](https://huggingface.co/Qwen/Qwen3-ASR-0.6B) works with the same hook and the same `transcribe` / `stream` API. It detects the language by itself; pass `language` (one of [`QWEN3_ASR_LANGUAGES`](../../06-api-reference/variables/QWEN3_ASR_LANGUAGES.md)) only to force one:
+
+```typescript
+const stt = useSpeechToText(models.speechToText.QWEN3_ASR.SIZE_0_6B.DEFAULT);
+
+const transcript = await stt.transcribe(audioData); // auto-detect
+const french = await stt.transcribe(audioData, { language: 'fr' });
+```
+
 ## Available Models
 
 The library provides ready-to-use Whisper models from the [Software Mansion HuggingFace Whisper Collection](https://huggingface.co/collections/software-mansion/whisper), available in [`models.speechToText`](../../06-api-reference/variables/models.md#speechtotext):
 
-| Model Family      | Variants                                                                                                                                                                  | Size Range         | Supported Backends                                            | Languages                                                                                              | Notes                                                           |
-| :---------------- | :------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | :----------------- | :------------------------------------------------------------ | :----------------------------------------------------------------------------------------------------- | :-------------------------------------------------------------- |
-| **Whisper Tiny**  | [`Multilingual`](../../06-api-reference/variables/models.md#speechtotextwhispertiny), [`English`](../../06-api-reference/variables/models.md#speechtotextwhisperentiny)   | 57.1 MB – 221.8 MB | XNNPACK (CPU), Core ML (Apple), MLX (Apple), Vulkan (Android) | English / [`WHISPER_LANGUAGES`](../../06-api-reference/variables/WHISPER_LANGUAGES.md) (99+ languages) | Ultra-fast transcription with minimal RAM usage.                |
-| **Whisper Base**  | [`Multilingual`](../../06-api-reference/variables/models.md#speechtotextwhisperbase), [`English`](../../06-api-reference/variables/models.md#speechtotextwhisperenbase)   | 97.8 MB – 380.2 MB | XNNPACK (CPU), Core ML (Apple), MLX (Apple), Vulkan (Android) | English / [`WHISPER_LANGUAGES`](../../06-api-reference/variables/WHISPER_LANGUAGES.md) (99+ languages) | Balanced accuracy and speed for general voice dictation.        |
-| **Whisper Small** | [`Multilingual`](../../06-api-reference/variables/models.md#speechtotextwhispersmall), [`English`](../../06-api-reference/variables/models.md#speechtotextwhisperensmall) | 276.0 MB – 1.05 GB | XNNPACK (CPU), Core ML (Apple), MLX (Apple), Vulkan (Android) | English / [`WHISPER_LANGUAGES`](../../06-api-reference/variables/WHISPER_LANGUAGES.md) (99+ languages) | High-capacity model for complex, noisy, or multi-speaker audio. |
+| Model Family       | Variants                                                                                                                                                                  | Size Range          | Supported Backends                                            | Languages                                                                                              | Notes                                                           |
+| :----------------- | :------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | :------------------ | :------------------------------------------------------------ | :----------------------------------------------------------------------------------------------------- | :-------------------------------------------------------------- |
+| **Whisper Tiny**   | [`Multilingual`](../../06-api-reference/variables/models.md#speechtotextwhispertiny), [`English`](../../06-api-reference/variables/models.md#speechtotextwhisperentiny)   | 57.1 MB – 221.8 MB  | XNNPACK (CPU), Core ML (Apple), MLX (Apple), Vulkan (Android) | English / [`WHISPER_LANGUAGES`](../../06-api-reference/variables/WHISPER_LANGUAGES.md) (99+ languages) | Ultra-fast transcription with minimal RAM usage.                |
+| **Whisper Base**   | [`Multilingual`](../../06-api-reference/variables/models.md#speechtotextwhisperbase), [`English`](../../06-api-reference/variables/models.md#speechtotextwhisperenbase)   | 97.8 MB – 380.2 MB  | XNNPACK (CPU), Core ML (Apple), MLX (Apple), Vulkan (Android) | English / [`WHISPER_LANGUAGES`](../../06-api-reference/variables/WHISPER_LANGUAGES.md) (99+ languages) | Balanced accuracy and speed for general voice dictation.        |
+| **Whisper Small**  | [`Multilingual`](../../06-api-reference/variables/models.md#speechtotextwhispersmall), [`English`](../../06-api-reference/variables/models.md#speechtotextwhisperensmall) | 276.0 MB – 1.05 GB  | XNNPACK (CPU), Core ML (Apple), MLX (Apple), Vulkan (Android) | English / [`WHISPER_LANGUAGES`](../../06-api-reference/variables/WHISPER_LANGUAGES.md) (99+ languages) | High-capacity model for complex, noisy, or multi-speaker audio. |
+| **Qwen3-ASR 0.6B** | [`0.6B`](../../06-api-reference/variables/models.md#speechtotextqwen3_asr)                                                                                                | 707.7 MB – 808.1 MB | XNNPACK (CPU), Core ML + MLX (Apple)                          | [`QWEN3_ASR_LANGUAGES`](../../06-api-reference/variables/QWEN3_ASR_LANGUAGES.md) (30 languages)        | Automatic language detection.                                   |
 
 ## API Reference
 
 ### Hooks & Pipelines
 
 - [`useSpeechToText()`](../../06-api-reference/functions/useSpeechToText.md) — React hook for Whisper model loading, downloading, and live transcription state.
 - [`createWhisperSpeechToText()`](../../06-api-reference/functions/createWhisperSpeechToText.md) — Imperative factory for Whisper Speech-to-Text pipelines.
+- [`createQwen3AsrSpeechToText()`](../../06-api-reference/functions/createQwen3AsrSpeechToText.md) — Imperative factory for Qwen3-ASR Speech-to-Text pipelines.
 
 ### Types & Options
 
@@ -226,15 +240,18 @@ The library provides ready-to-use Whisper models from the [Software Mansion Hugg
 - [`WhisperSttOptions`](../../06-api-reference/type-aliases/WhisperSttOptions.md) — Per-call transcription options (`language`).
 - [`WhisperStreamOptions`](../../06-api-reference/type-aliases/WhisperStreamOptions.md) — Live microphone streaming options (`language`, `vadOptions`).
 - [`WhisperLanguage`](../../06-api-reference/type-aliases/WhisperLanguage.md) — Union of supported Whisper language codes.
+- [`Qwen3AsrSttModel`](../../06-api-reference/type-aliases/Qwen3AsrSttModel.md), [`Qwen3AsrSttOptions`](../../06-api-reference/type-aliases/Qwen3AsrSttOptions.md), [`Qwen3AsrLanguage`](../../06-api-reference/type-aliases/Qwen3AsrLanguage.md) — Qwen3-ASR counterparts.
 
 ### Constants & Model Presets
 
 - [`WHISPER_SAMPLE_RATE_HZ`](../../06-api-reference/variables/WHISPER_SAMPLE_RATE_HZ.md) — Target audio sample rate expected by Whisper models (16000 Hz).
 - [`WHISPER_LANGUAGES`](../../06-api-reference/variables/WHISPER_LANGUAGES.md) — Array of 99+ supported language codes.
+- [`QWEN3_ASR_LANGUAGES`](../../06-api-reference/variables/QWEN3_ASR_LANGUAGES.md) — Language codes Qwen3-ASR can be forced to.
 - [`models.speechToText`](../../06-api-reference/variables/models.md#speechtotext) — Pre-configured Whisper models registry.
 
 :::info Source Code
 View the implementation on GitHub:
 
 - [`src/extensions/speech/tasks/whisperSpeechToText.ts` ↗](https://github.com/software-mansion/react-native-executorch/blob/main/packages/react-native-executorch/src/extensions/speech/tasks/whisperSpeechToText.ts)
+- [`src/extensions/speech/tasks/qwen3AsrSpeechToText.ts` ↗](https://github.com/software-mansion/react-native-executorch/blob/main/packages/react-native-executorch/src/extensions/speech/tasks/qwen3AsrSpeechToText.ts)
   :::
diff --git a/packages/react-native-executorch/__tests__/api/__snapshots__/apiSurface.test.ts.snap b/packages/react-native-executorch/__tests__/api/__snapshots__/apiSurface.test.ts.snap
@@ -125,6 +125,8 @@ exports[`public API surface matches the recorded export list 1`] = `
   "PASCAL_VOC_LABELS",
   "PRIVACY_FILTER_NEMOTRON_LABELS",
   "PRIVACY_FILTER_OPENAI_LABELS",
+  "QWEN3_ASR_LANGUAGES",
+  "QWEN3_ASR_SAMPLE_RATE_HZ",
   "RnExecuTorchError",
   "SUPERTONIC_DEFAULT_VOICE_NAMES",
   "SUPERTONIC_SAMPLE_RATE",
@@ -142,6 +144,7 @@ exports[`public API surface matches the recorded export list 1`] = `
   "createObjectDetector",
   "createPaddleOcr",
   "createPrivacyFilter",
+  "createQwen3AsrSpeechToText",
   "createResourceScope",
   "createSdxsTextToImage",
   "createSemanticSegmenter",

diff --git a/packages/react-native-executorch/__tests__/api/modelVariants.test.ts b/packages/react-native-executorch/__tests__/api/modelVariants.test.ts
@@ -320,6 +320,17 @@ describe('variant selection rules', () => {
     expect(defaultKeyOf(distiluse(registry))).toBe('XNNPACK_8DA4W');
   });
 
+  it('picks a variant naming two backends only when both are linked', () => {
+    // Qwen3-ASR's MLX file runs its audio tower on Core ML.
+    const qwen3Asr = (registry: Node): Node =>
+      ((registry.speechToText as Node).QWEN3_ASR as Node).SIZE_0_6B as Node;
+
+    expect(defaultKeyOf(qwen3Asr(registryFor({ os: 'ios' })))).toBe('MLX_COREML_INT4');
+    expect(
+      defaultKeyOf(qwen3Asr(registryFor({ os: 'ios', backends: ['XnnpackBackend', 'MLXBackend'] })))
+    ).toBe('XNNPACK_8DA4W');
+  });
+
   it('applies a pin only on the platform it names', () => {
     // The same group pins MLX for iOS only; Android resolves by its own order.
     expect(defaultKeyOf(distiluse(registryFor({ os: 'android' })))).toBe('VULKAN_FP16');
@@ -363,6 +374,13 @@ describe('feature map', () => {
       if (typeof value === 'string') {
         const match = value.match(/\/(xnnpack|coreml|mlx|vulkan)\//);
         if (match) found.add(match[1]!);
+        // A file can need a second backend, which only its name records:
+        // `mlx/qwen3_asr_0_6b_coreml_mlx_4w.pte` runs its audio tower on Core ML.
+        if (value.endsWith('.pte')) {
+          for (const part of value.split('/').pop()!.split('_')) {
+            if (/^(xnnpack|coreml|mlx|vulkan)$/.test(part)) found.add(part);
+          }
+        }
       } else if (Array.isArray(value)) value.forEach(walk);
       else if (isObject(value)) Object.values(value).forEach(walk);
     };