[msluszniak]

Description

• New createQwen3AsrSpeechToText task for Qwen3-ASR 0.6B, same transcribe/stream API as Whisper; language is auto-detected or forced via QWEN3_ASR_LANGUAGES.
• useSpeechToText accepts Qwen3-ASR models; registry entry models.speechToText.QWEN3_ASR.SIZE_0_6B (XNNPACK 8da4w; iOS default Core ML audio tower + MLX decoder; HF react-native-executorch-qwen-3-asr @ v0.11.0).
• A variant key naming two backends (MLX_COREML_INT4) is the default only when both are linked; loading a method whose backend is not linked now names it.
• Speech demo screens list the model; docs section added.

Introduces a breaking change?

• Yes
• No

Type of change

• Bug fix (change which fixes an issue)
• New feature (change which adds functionality)
• Documentation update (improves or adds clarity to existing documentation)
• Other (chores, tests, code style improvements etc.)

Tested on

• iOS
• Android

Testing instructions

• apps/speech: pick "Qwen3-ASR 0.6B (CPU)" on the audio file or microphone screen and transcribe.

Screenshots

Related issues

#769

Checklist

• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have updated the documentation accordingly
• My changes generate no new warnings

Additional notes

• WER on 100 LibriSpeech test-clean clips: 2.46% (fp32 2.14%). Decoder layer 0 must stay 8-bit or transcripts derail.
• iPhone 16: Core ML + MLX 293 ms per clip vs XNNPACK 468, WER 2.51%.
• No Vulkan: on a Galaxy S26 Ultra it wins prefill but loses encoder and decode (whole clip 451 vs 360 ms).
• Exporters and findings: export-scripts !41. ExecuTorch fixes found on the way: Vulkan: keep a reshape that follows a dynamically quantized linear pytorch/executorch#24055, #24056.

[msluszniak]

@barhanc I forgot that I need to rebuild libs for this one to cover all backends, will let you know when this will be ready since I want to batch this with Szymon's changes to MIGAN.