Fastest and only macOS Dictation app with on-device STT and custom trained AI enhancement model. Windows, iOS testflight! Android, Linux waitlist open! A local Wispr Flow alternative. Follow https://x.com/fluidvoiceapp for more
import AVFoundation
import CryptoKit
@testable import FluidVoice_Debug
import Foundation
import XCTest
#if arch(arm64)
import FluidAudio
#endif
@MainActor
final class ParakeetCompactModelSmokeTests: XCTestCase {
func testShippedRuntimeCheckRecordingIsBundledAndUnchanged() throws {
let url = try XCTUnwrap(CompactSpeechModelRuntimeCheck.recordingURL)
let data = try Data(contentsOf: url)
XCTAssertEqual(data.count, 131_162)
XCTAssertEqual(
SHA256.hash(data: data).map { String(format: "%02x", $0) }.joined(),
"961ac4a204b913fb953e1ef9a7d4a54a235f69a3bf0a818a6712968bacdde68d"
)
let samples = try CompactSpeechModelRuntimeCheck.samples()
XCTAssertEqual(samples.count, 65_581)
XCTAssertTrue(samples.allSatisfy(\.isFinite))
XCTAssertTrue(samples.contains { abs($0) > 0.01 })
}
func testRuntimeCheckRequiresFiveDistinctWholeWords() {
XCTAssertTrue(CompactSpeechModelRuntimeCheck.accepts("QUICK, brown fox jumps over."))
XCTAssertTrue(CompactSpeechModelRuntimeCheck.accepts("The quick brown fox jumps over the lazy dog near quiet water."))
XCTAssertFalse(CompactSpeechModelRuntimeCheck.accepts("quick brown fox jumps"))
XCTAssertFalse(CompactSpeechModelRuntimeCheck.accepts("quick quick quick quick quick"))
XCTAssertFalse(CompactSpeechModelRuntimeCheck.accepts("quickly brownish foxes jumping overhead"))
XCTAssertFalse(CompactSpeechModelRuntimeCheck.accepts(""))
}
func testRuntimeCheckRejectsInvalidAndOversizedPCM() throws {
for data in [Data(), Data([0]), Data(repeating: 0, count: 32_001), Data(repeating: 0, count: 480_002)] {
XCTAssertThrowsError(try CompactSpeechModelRuntimeCheck.decode(data)) { error in
XCTAssertEqual(error as? CompactSpeechModelRuntimeCheck.Failure, .invalidRecording)
}
}
var data = Data(repeating: 0, count: 32_000)
data[0] = 0xff
data[1] = 0x7f
data[2] = 0
data[3] = 0x80
let samples = try CompactSpeechModelRuntimeCheck.decode(data)
XCTAssertEqual(samples.count, 16_000)
XCTAssertEqual(samples[0], 32_767.0 / 32_768)
XCTAssertEqual(samples[1], -1)
}
func testRuntimeDeadlineJoinsCanceledProofCleanup() async throws {
let activity = RuntimeCheckActivity()
do {
try await CompactSpeechModelRuntimeCheck.withDeadline(.milliseconds(10)) {
do {
try await Task.sleep(for: .seconds(60))
} catch {
await activity.finish()
throw error
}
}
XCTFail("A blocked proof must time out")
} catch {
XCTAssertEqual(error as? CompactSpeechModelRuntimeCheck.Failure, .timedOut)
}
let finished = await activity.finished
XCTAssertTrue(finished, "The installer must not delete staged files while proof still uses them")
}
func testCanceledRuntimeProofJoinsCleanupWithoutReportingBadWeights() async throws {
let activity = RuntimeCheckActivity()
let (started, continuation) = AsyncStream<Void>.makeStream()
let task = Task {
try await CompactSpeechModelRuntimeCheck.withDeadline(.seconds(60)) {
continuation.yield(())
continuation.finish()
do {
try await Task.sleep(for: .seconds(60))
} catch {
await activity.finish()
throw error
}
}
}
for await _ in started { break }
task.cancel()
do {
try await task.value
XCTFail("Canceled verification must not publish weights")
} catch {
XCTAssertTrue(error is CancellationError)
}
let finished = await activity.finished
XCTAssertTrue(finished)
}
func testSuccessfulRuntimeProofCancelsDeadline() async throws {
let activity = RuntimeCheckActivity()
try await CompactSpeechModelRuntimeCheck.withDeadline(.seconds(60)) {
await activity.finish()
}
let finished = await activity.finished
XCTAssertTrue(finished)
}
func testHostedOldWeightsStayUsableUntilExplicitValidatedUpdate() async throws {
#if arch(arm64)
guard ProcessInfo.processInfo.environment["FLUIDVOICE_COMPACT_MODEL_AUDIO"] != nil else {
throw XCTSkip("Explicitly enable hosted compact model runtime checks")
}
let root = FileManager.default.temporaryDirectory.appendingPathComponent("FluidVoiceCompactUpdate-\(UUID().uuidString)", isDirectory: true)
defer { try? FileManager.default.removeItem(at: root) }
let selection = SettingsStore.shared.selectedSpeechModel
let fixtures: [(SettingsStore.SpeechModel, Int64, String)] = [
(.fluidParakeetMini, 254_126_592, "12109f89c80b959847ac80ae963714a00af08f029144c3e4a2a8c1641f528a36"),
(.fluidParakeetPico, 161_201_152, "afcdd76d4aba0328c97c858ee5c4c227ab5e23cc6885de5299d009c3671fe5d6"),
]
for (model, bytes, oldHash) in fixtures {
let latest = try XCTUnwrap(model.parakeetDescriptor)
let oldURL = try XCTUnwrap(URL(string: "https://models.fluidvoice.app/parakeet/fluid-\(latest.variant.rawValue)/1.0.0/\(latest.folderName).tar"))
let old = latest.replacingArchive(url: oldURL, sha256: oldHash, byteCount: bytes)
let directory = try await ParakeetArchiveDownloader.ensurePresent(descriptor: old, in: root)
let probe = SpeechModelInstallationSnapshot.Probe(modelID: model.id, kind: .parakeet(latest))
let before = try await SpeechModelInstallationSnapshot.scanResult([probe], modelsDirectory: root)
XCTAssertEqual(before.installedIDs, [model.id])
XCTAssertEqual(before.updateAvailableIDs, [model.id])
XCTAssertEqual(before.installedArchiveHashes[model.id], oldHash)
do {
let provider = FluidAudioProvider(modelOverride: model, configureWordBoosting: false)
provider.modelCacheRootForTesting = root
try await provider.prepare()
XCTAssertTrue(provider.isReady)
XCTAssertEqual(provider.pronunciationModelKeyForTesting, old.pronunciationModelKey)
XCTAssertEqual(latest.installedArchiveSHA256(at: directory), oldHash)
}
let updater = FluidAudioProvider(modelOverride: model, configureWordBoosting: false, updateCompactWeights: true)
updater.modelCacheRootForTesting = root
try await updater.prepare()
XCTAssertTrue(latest.artifactsAreComplete(at: directory))
let after = try await SpeechModelInstallationSnapshot.scanResult([probe], modelsDirectory: root)
XCTAssertEqual(after.installedIDs, [model.id])
XCTAssertTrue(after.updateAvailableIDs.isEmpty)
XCTAssertEqual(after.installedArchiveHashes[model.id], latest.archiveSHA256)
let provider = FluidAudioProvider(modelOverride: model, configureWordBoosting: false)
provider.modelCacheRootForTesting = root
try await provider.prepare()
XCTAssertTrue(provider.isReady)
XCTAssertEqual(provider.pronunciationModelKeyForTesting, latest.pronunciationModelKey)
XCTAssertEqual(SettingsStore.shared.selectedSpeechModel, selection)
}
#else
throw XCTSkip("Compact Parakeet models require Apple silicon")
#endif
}
func testHostedCompactModelsDownloadLoadTranscribeAndExtractDictionaryFrames() async throws {
#if arch(arm64)
guard let audioPath = ProcessInfo.processInfo.environment["FLUIDVOICE_COMPACT_MODEL_AUDIO"] else {
throw XCTSkip("Set FLUIDVOICE_COMPACT_MODEL_AUDIO to an English test recording for the hosted model check")
}
let verificationSamples = try CompactSpeechModelRuntimeCheck.samples()
let file = try AVAudioFile(forReading: URL(fileURLWithPath: audioPath), commonFormat: .pcmFormatFloat32, interleaved: false)
XCTAssertEqual(file.processingFormat.sampleRate, 16_000)
XCTAssertEqual(file.processingFormat.channelCount, 1)
let buffer = try XCTUnwrap(AVAudioPCMBuffer(pcmFormat: file.processingFormat, frameCapacity: AVAudioFrameCount(file.length)))
try file.read(into: buffer)
let channel = try XCTUnwrap(buffer.floatChannelData?[0])
let samples = Array(UnsafeBufferPointer(start: channel, count: Int(buffer.frameLength)))
XCTAssertFalse(samples.isEmpty)
XCTAssertLessThanOrEqual(samples.count, 240_000)
let root = FileManager.default.temporaryDirectory.appendingPathComponent("FluidVoiceCompactModels-\(UUID().uuidString)", isDirectory: true)
defer { try? FileManager.default.removeItem(at: root) }
let selection = SettingsStore.shared.selectedSpeechModel
let hosted: [(ParakeetSpeechModelCatalog.Descriptor, SettingsStore.SpeechModel)] = [
(ParakeetSpeechModelCatalog.mini, .fluidParakeetMini),
(ParakeetSpeechModelCatalog.pico, .fluidParakeetPico),
(ParakeetSpeechModelCatalog.pro, .fluidParakeetPro),
]
for (descriptor, model) in hosted {
// Exercise first-download publication through the real provider's staged proof.
do {
let provider = FluidAudioProvider(modelOverride: model, configureWordBoosting: false)
provider.modelCacheRootForTesting = root
try await provider.prepare()
XCTAssertTrue(provider.isReady)
XCTAssertFalse(provider.isWordBoostingActive)
}
let directory = descriptor.cacheDirectory(in: root)
XCTAssertTrue(descriptor.artifactsAreComplete(at: directory))
let models = try await AsrModels.loadLocalOnly(from: directory, version: descriptor.asrModelVersion)
XCTAssertEqual(models.version, descriptor.asrModelVersion)
// Pro's Mac pack keeps one encoder; Mini and Pico ship theirs in pieces.
XCTAssertEqual(models.splitEncoder != nil, descriptor.variant != .pro)
XCTAssertEqual(models.encoderOutputShape.count, 3)
let hiddenDimension = try XCTUnwrap(models.encoderOutputShape.dropFirst().first)
XCTAssertEqual(hiddenDimension, models.version.encoderHiddenSize)
let manager = AsrManager(config: .default)
do {
try await manager.initialize(models: models)
let verification = try await manager.transcribe(verificationSamples, source: .microphone)
XCTAssertTrue(CompactSpeechModelRuntimeCheck.accepts(verification.text), "Shipped sentence failed: \(verification.text)")
let result = try await manager.transcribe(samples, source: .microphone)
XCTAssertGreaterThanOrEqual(result.text.split(separator: " ").count, 3)
let embedding = try await manager.pronunciationEmbedding(audioSamples: samples, focalSampleRange: 0..<samples.count)
XCTAssertEqual(embedding.values.count, models.version.encoderHiddenSize)
XCTAssertTrue(embedding.values.allSatisfy(\.isFinite))
let frames = try await DictionaryTemporalEncoder.encode(samples, models: models)
XCTAssertTrue(frames.isValid)
XCTAssertEqual(frames.hiddenSize, models.version.encoderHiddenSize)
XCTAssertEqual(SettingsStore.shared.selectedSpeechModel, selection)
await manager.cleanup()
} catch {
await manager.cleanup()
throw error
}
}
#else
throw XCTSkip("Compact Parakeet models require Apple silicon")
#endif
}
}
private actor RuntimeCheckActivity {
private(set) var finished = false
func finish() { self.finished = true }
}