ailia MODELS Kotlin

Android Studio sample collection for running ailia SDK models from Kotlin. Each Ailia*Sample.kt keeps model download, preprocessing, inference, postprocessing, and resource release close together so one model can be copied into another application without bringing MainActivity with it.

Test environment

• macOS / Windows 11
• Android Studio 2025.1.3
• Gradle 8.13.2
• Kotlin 1.8.22
• minSdk 21 / targetSdk 35
• ailia SDK 1.5.0

Setup

Fetch the JNI modules, then open the repository root in Android Studio.

git submodule update --init --recursive

Models bundled in res/raw run without a download. Other models are downloaded when Run, Record, Generate, or Send is pressed and are stored in the application's durable model directory. A progress bar is shown while a transfer is active.

Supported models

Category	Model sample	Runtime
Pose Estimation	Lightweight Human Pose Estimation	ailia SDK
Object Detection	YOLOX TFLite, YOLOX ONNX, DETR ResNet50	ailia TFLite Runtime / ailia SDK
Object Tracking	ByteTrack with YOLOX	ailia Tracker + detector runtime
Image Classification	MobileNetV2 / ResNet50 TFLite, MobileNetV2 / ResNet50 / ViT-B/16 ONNX	ailia TFLite Runtime / ailia SDK
Background Removal	U-2-Net	ailia SDK
Zero-Shot Classification	multilingual-MiniLMv2 L12	ailia Tokenizer + ailia SDK
Speech to Text	Whisper / SenseVoice Small	ailia AI Speech
Speaker Verification	WeSpeaker ResNet34 (VoxCeleb) + Silero VAD v6	ailia SDK
Voice Filtering	VoiceFilter + dynamic d-vector embedder + Silero VAD v6	ailia SDK
Text to Speech	GPT-SoVITS V1 / V2 / V3 / V2-Pro / V2-Pro Distill JA	ailia AI Voice
LLM	Gemma 4 E2B / E4B / Gemma 2 2B	ailia LLM
Multimodal LLM	Gemma 3 4B	ailia LLM

Copying one sample into an application

Add only the JNI modules needed by the selected sample.

Sample	Required Gradle modules
ONNX vision, U-2-Net, DETR	ailia-sdk-jni
TFLite vision	ailia-tflite-jni
ByteTrack	ailia-tracker-jni plus the selected detector runtime
multilingual-MiniLMv2	ailia-sdk-jni, ailia-tokenizer-jni
Whisper / SenseVoice	ailia-speech-jni
WeSpeaker + Silero VAD v6	ailia-sdk-jni
VoiceFilter + Silero VAD v6	ailia-sdk-jni
GPT-SoVITS	ailia-voice-jni, ailia-audio-jni, ailia-sdk-jni
Gemma text or multimodal	ailia-llm-jni

For a downloadable model, also copy ModelDownloader.kt. It uses a temporary file, validates HTTP status and size, records a SHA-256 sidecar, and only then moves the model into place. The sample constructors require a model directory so storage ownership is explicit.

val modelDirectory = ModelDownloader.modelDirectory(context)
val sample = AiliaDetrSample(modelDirectory)

executor.execute {
    if (sample.downloadModel() && sample.initialize(environmentId)) {
        val inference = sample.detect(bitmap)
        val detections: List<DetectionResult> = inference?.value.orEmpty()
    }
}

Run native initialization, inference, and release on the same serial executor. Do not release a model from the UI thread while inference is active.

The inference-facing result classes are independent of the SDK wrappers:

• Detection and tracking: DetectionResult.kt, ModelInferenceResult.kt, CocoLabels.kt, and CategoryColors.kt
• Classification: ClassificationResult.kt, ModelInferenceResult.kt, and ImageNetLabels.kt
• Background removal: ModelInferenceResult.kt (SegmentationMask)
• Speaker verification: SpeakerVerificationAudio.kt and the session-only SpeakerProfileStore.kt
• Voice filtering: VoiceFilterAudio.kt and the session-only SpeakerProfileStore.kt

detect, classify, and predictMask return typed results without drawing. drawDetections and drawMask are optional Android renderers. The process... methods remain small compatibility wrappers used by the demo UI.

Models in res/raw must be copied with their sample. Downloaded filenames and URLs are declared beside each model enum or at the top of its sample class. INTERNET is required for those downloads; CAMERA and RECORD_AUDIO are requested only when the corresponding feature is used.

Screenshots

Pose Estimation	Object Detection	Tracking
Classification	Zero-Shot Classification	Speech to Text
Text to Speech	LLM	Multimodal LLM