v0.11.0c

v0.11.0c

Patch release focused on Android Whisper ASR GPU split execution and repeated-utterance latency.

Changed

• Added native compiled-model caching for Whisper ASR smoke execution.
  • The first utterance compiles the encoder/decoder models.
  • Later utterances in the same app process reuse the compiled models instead of recompiling every run.
  • ASR JSON/status output now reports compiledModelCache as miss or hit.
• Updated the Unity Android AAR with the cached Whisper ASR path.
• Updated ASR smoke tooling to support repeated benchmark runs through -BenchmarkRuns.
• Improved Whisper f32 encoder companion resolution so whisper_tiny_30s_f32_encoder.tflite is preferred while the legacy alias is still supported.
• Restored the patch-based Unity AAR build flow so the LiteRT-LM submodule can be patched and rebuilt from the Unity repository.

Verified

• Built the patched Android AAR successfully.
• Built the Android ASR smoke APK successfully.
• Ran whisper_tiny_30s_f32.tflite + whisper_tiny_30s_f32_encoder.tflite on the physical Qualcomm Android device with GPU_FP16.
• 10-run ASR result:
  • run 1: compiledModelCache=miss
  • runs 2-10: compiledModelCache=hit
  • backend: GPU_ENCODER_CPU_DECODER
  • average elapsed: 2.372s
  • average compile: 0.153s
  • average encode: 0.097s
  • average decode: 1.959s

Notes

• Whisper ASR GPU remains split execution: GPU encoder, CPU decoder.
• whisper_tiny_30s_i8.tflite remains the smallest CPU default for Korean/English ASR.
• whisper_tiny_30s_f32.tflite plus whisper_tiny_30s_f32_encoder.tflite is the recommended path for validating Whisper GPU ASR.

v0.11.0b

LiteRT-LM-Unity v0.11.0b

This release folds the ASR smoke-test checkpoint into main and refreshes the public documentation structure.

Added

• Android ASR smoke-test support for Whisper and Parakeet validation flows.
• Whisper Tiny Korean/English smoke coverage using whisper_tiny_30s_i8.tflite.
• Parakeet English smoke coverage using parakeet_tdt_0.6b_v3_5s_i8.tflite.
• Whisper Base reference support using whisper_base_30s_f32.tflite from litert-community/whisper-base.
• docs/asr-details.md with ASR setup, benchmark results, and smoke-test commands.
• docs/llm-details.md with LLM setup, benchmark results, and smoke-test notes.

Changed

• README now links to LLM 세부 설명 and ASR 세부 설명 instead of exposing benchmark-specific document names.
• README now documents that ASR models require a matching tokenizer under Assets/StreamingAssets.
• Detail documents are organized as setup, benchmarks, then smoke tests, with benchmark information placed before smoke-test commands.
• Older benchmark and handoff notes were moved under docs/.
• Android AAR patch and smoke build tooling were refreshed for the current ASR test path.

Notes

• Recommended ASR default: whisper_tiny_30s_i8.tflite for Korean/English Android CPU smoke testing.
• parakeet_tdt_0.6b_v3_5s_i8.tflite remains English-only.
• whisper_base_30s_f32.tflite is kept as a quality comparison model, not the default recommendation.
• External generated/export folders are intentionally not included in this release commit.

v0.11.0a - support parakeet

Full Changelog: v0.11.0...v0.11.0a

v0.11.0 - LiteRT-LM Unity bridge

LiteRT-LM Unity v0.11.0

This release aligns the Unity integration version with LiteRT-LM v0.11.0 and refreshes the Android bridge from a clean Docker/AAR reproduction.

Highlights

• Rebuilt Assets/Plugins/Android/litertlm-unity-bridge.aar from the LiteRT-LM submodule and Unity AAR patch workflow.
• Fixed clean AAR builds by copying root Bazel inputs that were previously omitted from the patched workspace:
  • requirements.txt
  • all root PATCH.* files referenced by WORKSPACE
• Kept all build scratch data inside the Unity workspace temp directory.
• Verified the generated AAR can be consumed by Unity Android builds.

Verification

• Removed project-specific Docker image/container state before reproducing the AAR build.
• Ran Tools/Windows/Build-LiteRtLmUnityAarFromPatch.ps1 -BazelJobs 8 successfully.
• Generated AAR: Builds/AndroidAar/litertlm-unity-bridge.aar (30,750,129 bytes).
• Copied AAR into Assets/Plugins/Android/litertlm-unity-bridge.aar.
• Built Unity Android smoke APK in batchmode with model.litertlm and GPU backend.
• Generated APK: Builds/Android/LiteRtLmAndroidSmokeTest-model-GPU.apk (93,568,021 bytes).
• Unity log confirmed Build Finished, Result: Success and LiteRT-LM Android AVD smoke APK built successfully.

Model Format Note

LiteRT-LM runtime code recognizes .task as a model format, and the Unity build scripts stage *.task files as StreamingAssets. The file still has to be LiteRT-compatible; non-compatible or metadata-incomplete task bundles can fail at runtime.

Build Notes

The Unity custom AAR patch is currently validated against the LiteRT-LM submodule used by this repository. Update the patch and rerun the clean AAR/APK checks when moving the submodule to a newer LiteRT-LM revision.

v0.1.0 - Initial Unity Editor test release

v0.1.0 - Initial Unity Editor test release

Clean first release of the LiteRT-LM Unity integration. The repository history is published as a single first commit and includes the Unity Editor validation path used to verify Windows Editor behavior.

Included

• Unity sample scene and controller for running LiteRT-LM from the Editor.
• Windows CLI fallback client for Unity Editor execution.
• PowerShell sample runner for litert_lm_main.windows_x86_64.exe.
• Batchmode Unity Editor self-test launcher.
• Conversation test scene with 10 prompts covering short prompts, Korean input, mixed-language prompts, longer diagnostic prompts, and context recall.
• Status-file and Unity Console debug logging for self-test progress.
• UTF-8 stdout/stderr handling for Korean text and emoji output.
• IME-aware IMGUI text fields for Korean prompt input.

Repository contents

• Assets/StreamingAssets/model.litertlm is committed as the small test model.
• Assets/StreamingAssets/model.litertlm.meta is committed for Unity asset stability.
• Other StreamingAssets files are ignored by default, including downloaded models, generated cache files, and runtime cache metadata.
• Generated .xnnpack_cache files are not part of this release.

Validation

Validated in Unity 6000.4.4f1 using the Windows Editor batchmode test path:

.\Tools\Windows\Run-LiteRtLmEditorSelfTest.ps1 `
  -MaxAttempts 1 `
  -ExecuteMethod 'LiteRTLM.Unity.Editor.LiteRtLmBuild.RunWindowsConversationSceneTestBatchmode' `
  -StatusRelativePath 'Builds\Logs\LiteRtLmConversationTest.status.txt' `
  -TestName 'Unity conversation scene test'

Result:

• Unity domain reload completed before the scene test.
• 10/10 conversation turns completed.
• Final context recall returned LRT-CTX-042.
• Unity process exit code was 0.

Notes

• The Windows Editor path starts the CLI process through the PowerShell wrapper for stability.
• Conversation continuity in the Editor fallback path is maintained by injecting summarized prior turns into each subsequent prompt.
• Large local models such as gemma-4-E2B-it.litertlm can be used for manual testing, but they are intentionally not committed.