Design Evolution and Best Practices for Android Audio Recorder in the FluentSpeak App

The FluentSpeak app has relied on a long‑standing Recorder class that records audio while scoring user speech, producing a score report and audio file.

Android Recorder Selection Android provides two Java‑level recorder APIs: MediaRecorder (produces encoded files) and AudioRecorder (provides raw PCM data). The app needs PCM for custom processing.

AudioFocus Handling When recording starts, any ongoing playback should be paused; therefore handling AudioFocus is recommended to avoid conflicts with other audio sources.

Basic AudioRecorder Usage Pattern 1. Create Recorder, Encoder, and Scorer on a dedicated Recorder thread. 2. Loop to read PCM data from AudioRecorder and feed it to Encoder and Scorer. 3. On external stop, set a run flag to false, exit the loop, close Encoder and Scorer, and save results. 4. Release all resources.

Problems Encountered - Encoder and Scorer processing blocks in the loop caused buffer overruns, leading to lost audio data and scoring errors. - Feature growth required supporting multiple audio formats and varied scoring logic, making the Recorder code increasingly complex. - Testing was hard because the Recorder depended on actual Android hardware; reproducing recordings reliably was difficult. - Native NDK components (Encoder/Scorer) could crash, risking the whole app.

Solutions Implemented - Added a ProcessThread so Encoder and Scorer run separately, preventing the Recorder thread from being blocked. - Introduced an AudioProcessor abstraction to unify Encoder and Scorer handling; the Recorder now delegates processing to injected AudioProcessor instances. - Created WavFileRecorder to simulate recordings from files, enabling fast, repeatable tests without hardware. - Developed IAudioProcessorService (AIDL) to run AudioProcessor in a separate Service process; native crashes only kill the service process, leaving the main app stable.

Summary of Practices 1. Process PCM data on a separate thread to avoid data loss. 2. Abstract processing logic into AudioProcessor for modularity. 3. Use WavFileRecorder for file‑based testing. 4. Run AudioProcessor in a separate Service via AIDL to isolate native crashes.

The resulting implementation is packaged as LingoRecorder , which demonstrates these patterns.