{
  "project": "Parakey",
  "benchmark": "swift-bench",
  "source": "https://github.com/rcourtman/parakey/tree/main/experiments/swift-bench",
  "environment": {
    "machine": "Mac mini M4",
    "cpu_cores": 10,
    "memory_gb": 16,
    "macos": "26.4.1",
    "swift": "6.3"
  },
  "methodology": {
    "trials_per_backend_per_clip": 5,
    "reported_statistic": "p50",
    "warmup": "First inference after model load excluded from p50",
    "audio_format": "16 kHz mono Float32 WAV produced with experiments/swift-bench/generate-test-audio.sh"
  },
  "backends": {
    "fluid": {
      "label": "FluidAudio Swift SDK, Parakeet TDT 0.6B v3, CoreML",
      "hardware": "Apple Neural Engine"
    },
    "parakey_mlx": {
      "label": "Prior parakeet_mlx path through MLX",
      "hardware": "GPU"
    }
  },
  "results": [
    {
      "clip": "short-clean",
      "duration_seconds": 2.50,
      "fluid_p50_ms": 92.4,
      "parakey_mlx_p50_ms": 145.4,
      "speed_ratio": 1.57
    },
    {
      "clip": "medium-clean",
      "duration_seconds": 3.99,
      "fluid_p50_ms": 96.1,
      "parakey_mlx_p50_ms": 176.3,
      "speed_ratio": 1.83
    },
    {
      "clip": "disfluent",
      "duration_seconds": 5.31,
      "fluid_p50_ms": 94.1,
      "parakey_mlx_p50_ms": 185.9,
      "speed_ratio": 1.97
    },
    {
      "clip": "longer-technical",
      "duration_seconds": 9.49,
      "fluid_p50_ms": 152.4,
      "parakey_mlx_p50_ms": 300.9,
      "speed_ratio": 1.97
    }
  ],
  "notes": [
    "The benchmark measures compute latency, not energy use.",
    "Both tested backends produced essentially identical transcripts on the synthetic audio.",
    "Apple SpeechAnalyzer wiring exists in the benchmark but is currently blocked by a CLI Info.plist and entitlement gap."
  ]
}