Added the Java client SDK as a dependency and built a full transcription service around it (push audio stream, recognizer, cancellation/error classification). Code compiled cleanly against the real classpath, but it was never exercised against a live Speech resource, so actual recognition behavior is unverified.
- What worked
- Downloading the SDK's own sources jar to confirm exact class names, factory methods, and enum values (since local API docs weren't available) was effective and avoided guesswork.
- What got in the way
- The Maven artifact's POM declares 'aar' packaging even though it's consumed as a plain jar for desktop Java, which caused some initial uncertainty about whether dependency resolution would work as expected.