feat: migrate on-device Qwen3 ASR to CoreML for background Flow dictation

Replace MLX GPU inference with CoreML bundles so transcription continues
while the host app is backgrounded. Adds model download and warm-up,
vendored Qwen3Speech, and updates onboarding, settings, and copy for the
~1.6 GB CoreML package (iOS 18+).
This commit is contained in:
Rocky
2026-06-23 00:46:58 +08:00
parent 5e5122f172
commit df1c5ff32c
160 changed files with 22080 additions and 492 deletions
@@ -0,0 +1,42 @@
// UtteranceStreamChunkerTests.swift
// OSGKeyboardTests
import XCTest
@testable import OSGKeyboardShared
final class UtteranceStreamChunkerTests: XCTestCase {
private let config = FlowUtteranceChunkConfig(
maxChunkDurationSeconds: 1,
overlapDurationSeconds: 0.1,
pauseExtensionMaxSeconds: 0.2,
pauseRMSThreshold: 0.02,
sampleRate: 1_000
)
func testPauseAwareSplitPrefersSilenceNearWindowEnd() {
var buffer = [Float](repeating: 0.2, count: 900)
buffer.append(contentsOf: [Float](repeating: 0.001, count: 50))
buffer.append(contentsOf: [Float](repeating: 0.2, count: 100))
let split = UtteranceStreamChunker.pauseAwareSplitIndex(in: buffer, config: config)
XCTAssertGreaterThanOrEqual(split, config.maxChunkSamples)
XCTAssertLessThanOrEqual(split, config.maxChunkSamples + config.pauseExtensionSamples)
}
func testChunksEmitMultipleSegmentsForLongStream() async {
let sampleCount = config.maxChunkSamples * 2 + 100
let samples = [Float](repeating: 0.05, count: sampleCount)
let (stream, continuation) = AsyncStream<AudioBufferSnapshot>.makeStream()
continuation.yield(AudioBufferSnapshot(samples: samples, sampleRate: Double(config.sampleRate)))
continuation.finish()
var received: [UtteranceAudioChunk] = []
for await chunk in UtteranceStreamChunker.chunks(from: stream, config: config) {
received.append(chunk)
}
XCTAssertGreaterThanOrEqual(received.count, 2)
XCTAssertTrue(received.last?.isLast == true)
}
}