[Alex-Wengg]

Fixes #4.

Problem

• feedStreamingAudio discarded the audio it was given and still resolved success: true. Streaming only worked through the microphone tap that startStreamingAsr installs itself.
• The podspec depended on FluidAudio ~> 0.7. CocoaPods trunk stops at 0.7.8 (Nov 2025) and goes read-only on December 2, 2026. 0.7.8 predates FluidAudio's current streaming managers, and the StreamingAsrManager class this module used was removed upstream.

Change

• Dependency: FluidAudio comes from Swift Package Manager through React Native's spm_dependency (FluidAudio 0.18.0 or later, up to the next minor). The React Native peer dependency is now 0.75 or later, and the podspec stops with a clear error on older versions. pod 'FluidAudio' is removed from the example Podfile.
• Port to FluidAudio 0.18: batch ASR, plus streaming through StreamingModelVariant.createManager(). The default model is parakeet-eou-320ms, the one FluidAudio's own iOS streaming demos use; JS can choose another with the new model option.
• feedStreamingAudio works: base64 PCM16 is decoded and fed to the manager. A new source: 'manual' starts a session without the microphone. Feeding a microphone session now rejects instead of silently dropping audio. The microphone tap and fed audio go through one ordered queue, and stop drains it before finishing.
• Example app: a "Stream Sample File" button feeds a bundled 16 kHz WAV (FluidAudio's Apache-2.0 test fixture) in 320 ms chunks. It also adds the missing NSMicrophoneUsageDescription, without which iOS kills the microphone demo on a device.

Breaking changes for JS users

• React Native 0.75 or later is required.
• Feeding audio needs start({ source: 'manual' }). source: 'system' is removed.
• StreamingUpdate.volatile is always empty; the transcript is in confirmed and text.
• start() and feed now return { success, source, model } and { success, samples }.

Verification

• npx tsc --noEmit passes. npx jest: 54/54.

• pod install in example/ios resolves FluidAudio 0.18.0.

• xcodebuild builds the example for generic/platform=iOS and generic/platform=iOS Simulator.

• The manual-feed path was driven on macOS with real speech. A harness ran this module's own AudioUtils.pcm16MonoBuffer (copied verbatim), the same ordered AsyncStream → appendAudio / processBufferedAudio → finish() session loop, and parakeet-eou-320ms from the FluidAudio 0.18.0 checkout the iOS build resolved. It fed base64 PCM16 in 320 ms chunks:

Clip	This PR	main (audio discarded)
example/assets/speech-16k-mono.wav, 16.9 s	"diff means comparing what is net new so what is the difference between the old code and the net new code and analyzing if there's any bugs or problems in that difference" (25 partial updates)	empty, 0 samples fed
FLEURS en_us_0003, 5.8 s	"the cabbage juice changed his color depending on how acidic basic alkaline the chemical is" (reference: "…changes color depending on how acidic or basic alkaline…")	empty

Not verified: model download, the transcript and microphone streaming inside the app on a device or simulator. No iOS Simulator runtime is installed on the build machine.

🤖 Generated with Claude Code