[michalharakal]

What

The streaming Moonshine runtime reran the audio frontend over the whole 68-frame window at every hop, and once more over the tail at finish(). That tail is most of the end-of-utterance flush.

With an optional ninth model file, frontend_step.vmfb, next to frontend.vmfb, three things change:

• Frontend: it runs in 0.32 s chunks (16 frames + 4 frames of causal left context) as the audio arrives, and its features are cached.
• Windows: they run only the encoder and adapter on the cached features.
• Padding: frames more than 4 frames past the audio can only see zero padding. They take the frontend's constant silence feature, captured at warm-up, instead of another graph call.

frontend_step.vmfb is the same frontend exported with MOONSHINE_FE_SAMPLES=6400. Without the file, behaviour is unchanged.

Also in this PR

• moonshine_stream_check.c: drives the library on a device without an app (fake JNIEnv, like whisper_session_check.c).
• stats(): gains feStepCalls / feStepMs.
• IreeMoonshineStream.OPTIONAL_FILE_NAMES: names the optional file.
• build-moonshine-stream.sh: always passes CMAKE_C_FLAGS. Before, a cached GEOMETRY from the previous build leaked into the next one.
• Libraries: both ABIs rebuilt from this source.

Evidence

Measured on an arm32 Mali device (Vulkan, IREE 3.11) with 40 German commands fed like a microphone (40 ms chunks, real-time pace). The 0.62.2 library ran 3 times and this branch ran 2 times:

	0.62.2	this branch
flush p50 / p95	319 / 324 ms	172 / 236 ms
finish p50	882 ms	737 ms
end of audio → final text p50	1.24 s	1.02–1.05 s
transcripts		identical, 40/40

• Speed: faster in 40/40 recordings, with a paired median of −147 ms on flush and −275 ms from end of audio to final text.
• Features: a build with -DMOONSHINE_FE_VERIFY also runs the whole-window frontend and compares its output. It found maxdiff 0 in 141/141 windows, so the features are bit-identical, including the silence-feature shortcut.
• Re-check: the committed armeabi-v7a library was re-checked on the device. Its transcripts match 0.62.2 in 40/40 recordings, with flush p50 172 ms.
• Not changed: the exact final re-decode (~565 ms p50).

Follow-ups (not in this PR)

• The blueprint asr-moonshine-v2-streaming-iree should export and compile frontend_step.vmfb.
• Materializers should copy OPTIONAL_FILE_NAMES when present.

🤖 Generated with Claude Code