{
  "project": "Presspeech",
  "benchmark": "swift-bench",
  "source": "https://github.com/rcourtman/presspeech/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"
    },
    "unified": {
      "label": "Candidate only: FluidAudio Swift SDK, Parakeet Unified 0.6B English offline batch, CoreML",
      "hardware": "Apple Neural Engine",
      "candidate_padding_ms": 250
    },
    "apple": {
      "label": "Candidate only: Apple SpeechAnalyzer + DictationTranscriber",
      "hardware": "Apple Neural Engine"
    },
    "nemotron_en": {
      "label": "Candidate only: FluidAudio Swift SDK, Nemotron Speech Streaming English 0.6B 1120 ms tier, CoreML",
      "hardware": "Apple Neural Engine"
    },
    "presspeech_mlx": {
      "label": "Prior parakeet_mlx path through MLX",
      "hardware": "GPU"
    }
  },
  "results": [
    {
      "clip": "short-clean",
      "duration_seconds": 2.50,
      "fluid_p50_ms": 92.4,
      "presspeech_mlx_p50_ms": 145.4,
      "speed_ratio": 1.57
    },
    {
      "clip": "medium-clean",
      "duration_seconds": 3.99,
      "fluid_p50_ms": 96.1,
      "presspeech_mlx_p50_ms": 176.3,
      "speed_ratio": 1.83
    },
    {
      "clip": "disfluent",
      "duration_seconds": 5.31,
      "fluid_p50_ms": 94.1,
      "presspeech_mlx_p50_ms": 185.9,
      "speed_ratio": 1.97
    },
    {
      "clip": "longer-technical",
      "duration_seconds": 9.49,
      "fluid_p50_ms": 152.4,
      "presspeech_mlx_p50_ms": 300.9,
      "speed_ratio": 1.97
    }
  ],
  "model_decision_checks": {
    "tail_word_regression": {
      "script": "experiments/swift-bench/run-tail-word-regression.sh",
      "purpose": "Synthetic final-word retention regression for short push-to-talk clips",
      "candidate_unified_padding_ms": 250
    },
    "private_real_dictation_comparison": {
      "script": "experiments/swift-bench/run-real-model-comparison.sh",
      "purpose": "Local-only v3-vs-Unified candidate comparison over private human dictation fixtures",
      "reported_metrics": [
        "average WER",
        "worst WER",
        "final-word failures",
        "p50 latency"
      ],
      "redacted_by_default": true
    },
    "public_candidate_smoke_2026_06_23": {
      "script": "experiments/swift-bench/run-real-dictation-regression.sh",
      "purpose": "Public 5-clip smoke comparison for production v3, Apple SpeechAnalyzer, and Nemotron English",
      "trials_per_clip": 3,
      "results": [
        {
          "backend": "v3",
          "avg_wer_percent": 1.4,
          "worst_wer_percent": 7.0,
          "final_word_failures": 0,
          "avg_p50_ms": 100.7
        },
        {
          "backend": "apple",
          "avg_wer_percent": 11.7,
          "worst_wer_percent": 18.2,
          "final_word_failures": 0,
          "avg_p50_ms": 384.8
        },
        {
          "backend": "nemotron-en",
          "avg_wer_percent": 5.3,
          "worst_wer_percent": 11.3,
          "final_word_failures": 0,
          "avg_p50_ms": 1032.7,
          "runtime_warning": "CoreML E5RT shape-inference warning emitted on each public clip"
        }
      ]
    },
    "release_gate": {
      "script": "experiments/swift-bench/run-release-asr-checks.sh",
      "runs_private_real_audio_when_present": true,
      "candidate_models_require_flag": "--include-candidate-models"
    }
  },
  "notes": [
    "The benchmark measures compute latency, not energy use.",
    "Both tested backends produced essentially identical transcripts on the synthetic audio.",
    "Unified English is candidate-only and is not exposed in the app unless future benchmarks and real dictation tests prove it improves on v3."
  ]
}