Whisper on Android: two paths to offline speech-to-text

What it does

This repo ships two Android apps and a Python conversion script to run OpenAI Whisper speech recognition entirely on-device. One app targets Java developers via the TensorFlow Lite Java API; the other uses the TensorFlow Lite Native API for better performance. Both handle 16 kHz mono 16-bit WAV audio and include a Recorder class for live microphone input. Pre-built APKs and pre-generated TFLite models are included if you want to skip the conversion step.

The interesting bit

The project is essentially a packaging layer around Niranjan Yadla’s earlier whisper.tflite port, but the dual-app structure is the useful part: you can compare Java vs. native inference overhead in the same codebase without writing your own JNI glue. The Python script in models_and_scripts handles the ONNX-to-TFLite conversion so you aren’t stuck with a single model size.

Key highlights

• Fully offline — no network required for transcription
• Two API surfaces: Java (whisper_java) and native C++ (whisper_native)
• Live transcription support via Recorder → writeBuffer() pipeline
• Pre-generated TFLite models and APKs for immediate testing
• Multilingual mode toggle via boolean flag in loadModel()

Caveats

• The README warns that “careful synchronization and error handling” are needed for smooth audio processing, which reads as “this is demo-grade, not production-polished”
• No benchmark numbers or latency claims are provided for either Java or native paths
• The project appears to be a packaging/integration effort rather than original model work; the heavy lifting is credited to Yadla’s upstream whisper.tflite

Verdict

Grab this if you need a working starting point for Whisper on Android and want to evaluate Java vs. native TFLite without building from scratch. Skip it if you need battle-tested, production-ready audio pipelines or detailed performance comparisons — you’ll be doing that profiling yourself.