Skip to content

README.md

Latest commit

 

History

History
36 lines (26 loc) · 2.04 KB

README.md

File metadata and controls

36 lines (26 loc) · 2.04 KB

LiteRT-LM-Unity

Unity integration sample for LiteRT-LM on Windows Editor and Android.

Requirements

  • Unity 6000.4.6f1
  • Windows and PowerShell for Editor and Android build scripts
  • Unity Android Build Support with SDK/NDK
  • Android device with adb access for hardware tests
  • Docker Desktop and Git for Windows Bash only when rebuilding the custom Android AAR

Recommended LLM Models

Rank Model Source Note
1 gemma-4-E2B-it.litertlm litert-community/gemma-4-E2B-it-litert-lm Primary quality model when memory allows.
2 Qwen2.5-0.5B-Instruct-q8.litertlm litert-community/Qwen2.5-0.5B-Instruct Fast CPU fallback for smaller memory and low latency.
3 gemma3-1b-it-int4.litertlm litert-community/Gemma3-1B-IT Compact Android native-GPU-capable fallback.

Recommended ASR Models

ASR models need a matching tokenizer directory under Assets/StreamingAssets. For example, Whisper Tiny uses whisper-tiny/tokenizer.json, and Whisper Base uses whisper-base/tokenizer.json. Use the tokenizer source links below when restoring those tokenizer folders.

Rank Model Model source Tokenizer source Note
1 whisper_tiny_30s_i8.tflite litert-community/whisper-tiny openai/whisper-tiny Recommended Korean/English ASR model for Android CPU use.
2 whisper_base_30s_f32.tflite litert-community/whisper-base openai/whisper-base Quality comparison only; it is large and not the default recommendation.

Details