Upload 205 files

manifest.json

@@ -1,376 +1,1371 @@
 {
-  "name": "nemotron-asr-streaming-multilingual-0.6b-coreml",
-  "version": "1.0.0",
+  "name": "Nemotron 3.5 ASR Streaming Multilingual 0.6B — CoreML",
   "base_model": "nvidia/nemotron-asr-streaming-multilingual-0.6b",
-  "base_model_class": "nemo.collections.asr.models.rnnt_bpe_models_prompt.EncDecRNNTBPEModelWithPrompt",
-  "framework": "coreml",
-  "deployment_target": {
-    "ios": "17.0",
-    "macos": "14.0"
-  },
-  "quantization": {
-    "encoder": "int8",
-    "preprocessor": "fp16",
-    "decoder": "fp16",
-    "joint": "fp16"
-  },
-  "model": {
-    "sample_rate": 16000,
-    "mel_features": 128,
-    "vocab_size": 13087,
-    "blank_idx": 13087,
-    "encoder_dim": 1024,
-    "decoder_hidden": 640,
-    "decoder_layers": 2,
-    "num_prompts": 128,
-    "default_prompt_id": 101
-  },
-  "loader": {
-    "library": "fluidaudio",
-    "class": "StreamingNemotronMultilingualAsrManager",
-    "preferred_format": "mlmodelc",
-    "fallback_format": "mlpackage"
-  },
-  "variants": [
-    {
-      "name": "80ms",
-      "chunk_ms": 80,
-      "chunk_mel_frames": 8,
-      "pre_encode_cache": 9,
-      "total_mel_frames": 17,
-      "att_context_size": [
-        56,
-        0
+  "architecture": "Conformer encoder + RNN-T decoder",
+  "runtime": "CoreML / Apple Neural Engine",
+  "benchmark_machine": "Apple M5 Pro / macOS 26.5",
+  "recipe": "LAYERPOS[42,13] mixed-precision encoder + per-language vocab pruning + B1 fusion + triple-stage pipelining + smart-spec K=4",
+  "no_retraining": true,
+  "no_calibration": true,
+  "tiers_ms": [
+    560,
+    1120,
+    2240,
+    4480
+  ],
+  "recommended_tier_ms": 2240,
+  "models": [
+    "de",
+    "en",
+    "es",
+    "fr",
+    "it",
+    "ja",
+    "multilingual",
+    "pt",
+    "zh"
+  ],
+  "ship_count": 36,
+  "notes": [
+    "2240ms (2s) is the RTFx-per-latency sweet spot for every model (+20-44% over 1120ms, quality-neutral).",
+    "Multilingual peaks at 2240ms (74.6 RTFx); its 4480ms tier craters (19.4) because the 13088-vocab joint exceeds ANE working-set.",
+    "560ms is the lowest-latency tier but off the trained 14-frame attention tiling: lower RTFx (~57) and a small quality cost. en measured (57.2 RTFx); other 560ms ships shipped unbenched.",
+    "Portuguese 2240ms WER is +3.1pp vs 1120ms (B1-fallback path, chunk-sensitive).",
+    "de/zh/ja keep-sets were derived from FLEURS-test transcripts: in-domain/optimistic numbers + OOV risk on out-of-domain text. Rebuild keep-set from a broader corpus for production.",
+    "Each <lang>/<tier>ms/ dir is a self-contained FluidAudio model bundle (point --model-dir at it)."
+  ],
+  "ships": [
+    {
+      "path": "en/560ms",
+      "language": "English",
+      "language_code": "en-US",
+      "chunk_ms": 560,
+      "latency_s": 0.56,
+      "chunk_mel_frames": 56,
+      "total_mel_frames": 65,
+      "att_context": [
+        42,
+        13
       ],
-      "cache_channel_shape": [
-        1,
-        24,
-        56,
-        1024
-      ],
-      "cache_time_shape": [
-        1,
-        24,
-        1024,
-        8
-      ],
-      "components": {
-        "preprocessor": {
-          "mlpackage": {
-            "path": "80ms/preprocessor.mlpackage",
-            "size_bytes": 608879
-          },
-          "mlmodelc": {
-            "path": "80ms/preprocessor.mlmodelc",
-            "size_bytes": 617594
-          }
-        },
-        "encoder": {
-          "mlpackage": {
-            "path": "80ms/encoder.mlpackage",
-            "size_bytes": 593323585
-          },
-          "mlmodelc": {
-            "path": "80ms/encoder.mlmodelc",
-            "size_bytes": 593506646
-          }
-        },
-        "decoder": {
-          "mlpackage": {
-            "path": "80ms/decoder.mlpackage",
-            "size_bytes": 29881569
-          },
-          "mlmodelc": {
-            "path": "80ms/decoder.mlmodelc",
-            "size_bytes": 29889159
-          }
-        },
-        "joint": {
-          "mlpackage": {
-            "path": "80ms/joint.mlpackage",
-            "size_bytes": 18916847
-          },
-          "mlmodelc": {
-            "path": "80ms/joint.mlmodelc",
-            "size_bytes": 18923548
-          }
-        }
-      },
-      "metadata_path": "80ms/metadata.json",
-      "tokenizer_path": "80ms/tokenizer.json",
-      "total_size_bytes": 1285964002
-    },
-    {
-      "name": "160ms",
-      "chunk_ms": 160,
-      "chunk_mel_frames": 16,
-      "pre_encode_cache": 9,
-      "total_mel_frames": 25,
-      "att_context_size": [
-        56,
-        0
+      "vocab_size": 988,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
       ],
-      "cache_channel_shape": [
-        1,
-        24,
-        56,
-        1024
-      ],
-      "cache_time_shape": [
-        1,
-        24,
-        1024,
-        8
-      ],
-      "components": {
-        "preprocessor": {
-          "mlpackage": {
-            "path": "160ms/preprocessor.mlpackage",
-            "size_bytes": 608879
-          },
-          "mlmodelc": {
-            "path": "160ms/preprocessor.mlmodelc",
-            "size_bytes": 611446
-          }
-        },
-        "encoder": {
-          "mlpackage": {
-            "path": "160ms/encoder.mlpackage",
-            "size_bytes": 593373136
-          },
-          "mlmodelc": {
-            "path": "160ms/encoder.mlmodelc",
-            "size_bytes": 593550818
-          }
-        },
-        "decoder": {
-          "mlpackage": {
-            "path": "160ms/decoder.mlpackage",
-            "size_bytes": 29881569
-          },
-          "mlmodelc": {
-            "path": "160ms/decoder.mlmodelc",
-            "size_bytes": 29883011
-          }
-        },
-        "joint": {
-          "mlpackage": {
-            "path": "160ms/joint.mlpackage",
-            "size_bytes": 18916847
-          },
-          "mlmodelc": {
-            "path": "160ms/joint.mlmodelc",
-            "size_bytes": 18917400
-          }
-        }
-      },
-      "metadata_path": "160ms/metadata.json",
-      "tokenizer_path": "160ms/tokenizer.json",
-      "total_size_bytes": 1286031086
-    },
-    {
-      "name": "320ms",
-      "chunk_ms": 320,
-      "chunk_mel_frames": 32,
-      "pre_encode_cache": 9,
-      "total_mel_frames": 41,
-      "att_context_size": [
-        56,
-        0
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 57.2,
+        "wer_pct": 4.1,
+        "cer_pct": 1.5,
+        "metric": "WER",
+        "n": 2620,
+        "test_set": "LibriSpeech test-clean",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "en/1120ms",
+      "language": "English",
+      "language_code": "en-US",
+      "chunk_ms": 1120,
+      "latency_s": 1.12,
+      "chunk_mel_frames": 112,
+      "total_mel_frames": 121,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 988,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
       ],
-      "cache_channel_shape": [
-        1,
-        24,
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 101.4,
+        "wer_pct": 3.81,
+        "cer_pct": 1.4,
+        "metric": "WER",
+        "n": 2620,
+        "test_set": "LibriSpeech test-clean",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "en/2240ms",
+      "language": "English",
+      "language_code": "en-US",
+      "chunk_ms": 2240,
+      "latency_s": 2.24,
+      "chunk_mel_frames": 224,
+      "total_mel_frames": 233,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 988,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 134.7,
+        "wer_pct": 3.7,
+        "cer_pct": 1.4,
+        "metric": "WER",
+        "n": 2620,
+        "test_set": "LibriSpeech test-clean",
+        "benched": true
+      },
+      "recommended": true
+    },
+    {
+      "path": "en/4480ms",
+      "language": "English",
+      "language_code": "en-US",
+      "chunk_ms": 4480,
+      "latency_s": 4.48,
+      "chunk_mel_frames": 448,
+      "total_mel_frames": 457,
+      "att_context": [
         56,
-        1024
-      ],
-      "cache_time_shape": [
-        1,
-        24,
-        1024,
-        8
-      ],
-      "components": {
-        "preprocessor": {
-          "mlpackage": {
-            "path": "320ms/preprocessor.mlpackage",
-            "size_bytes": 608879
-          },
-          "mlmodelc": {
-            "path": "320ms/preprocessor.mlmodelc",
-            "size_bytes": 617594
-          }
-        },
-        "encoder": {
-          "mlpackage": {
-            "path": "320ms/encoder.mlpackage",
-            "size_bytes": 593471709
-          },
-          "mlmodelc": {
-            "path": "320ms/encoder.mlmodelc",
-            "size_bytes": 593656949
-          }
-        },
-        "decoder": {
-          "mlpackage": {
-            "path": "320ms/decoder.mlpackage",
-            "size_bytes": 29881569
-          },
-          "mlmodelc": {
-            "path": "320ms/decoder.mlmodelc",
-            "size_bytes": 29889159
-          }
-        },
-        "joint": {
-          "mlpackage": {
-            "path": "320ms/joint.mlpackage",
-            "size_bytes": 18916847
-          },
-          "mlmodelc": {
-            "path": "320ms/joint.mlmodelc",
-            "size_bytes": 18923548
-          }
-        }
-      },
-      "metadata_path": "320ms/metadata.json",
-      "tokenizer_path": "320ms/tokenizer.json",
-      "total_size_bytes": 1286262430
-    },
-    {
-      "name": "560ms",
+        13
+      ],
+      "vocab_size": 989,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 136.7,
+        "wer_pct": 3.7,
+        "cer_pct": 1.4,
+        "metric": "WER",
+        "n": 2620,
+        "test_set": "LibriSpeech test-clean",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "es/560ms",
+      "language": "Spanish",
+      "language_code": "es-ES",
       "chunk_ms": 560,
+      "latency_s": 0.56,
       "chunk_mel_frames": 56,
-      "pre_encode_cache": 9,
       "total_mel_frames": 65,
-      "att_context_size": [
-        56,
-        0
+      "att_context": [
+        42,
+        13
       ],
-      "cache_channel_shape": [
-        1,
-        24,
-        56,
-        1024
-      ],
-      "cache_time_shape": [
-        1,
-        24,
-        1024,
-        8
-      ],
-      "components": {
-        "preprocessor": {
-          "mlpackage": {
-            "path": "560ms/preprocessor.mlpackage",
-            "size_bytes": 608879
-          },
-          "mlmodelc": {
-            "path": "560ms/preprocessor.mlmodelc",
-            "size_bytes": 617594
-          }
-        },
-        "encoder": {
-          "mlpackage": {
-            "path": "560ms/encoder.mlpackage",
-            "size_bytes": 593619582
-          },
-          "mlmodelc": {
-            "path": "560ms/encoder.mlmodelc",
-            "size_bytes": 593807012
-          }
-        },
-        "decoder": {
-          "mlpackage": {
-            "path": "560ms/decoder.mlpackage",
-            "size_bytes": 29881569
-          },
-          "mlmodelc": {
-            "path": "560ms/decoder.mlmodelc",
-            "size_bytes": 29889159
-          }
-        },
-        "joint": {
-          "mlpackage": {
-            "path": "560ms/joint.mlpackage",
-            "size_bytes": 18916847
-          },
-          "mlmodelc": {
-            "path": "560ms/joint.mlmodelc",
-            "size_bytes": 18923548
-          }
-        }
-      },
-      "metadata_path": "560ms/metadata.json",
-      "tokenizer_path": "560ms/tokenizer.json",
-      "total_size_bytes": 1286560366
-    },
-    {
-      "name": "1120ms",
+      "vocab_size": 831,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": null,
+        "wer_pct": null,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 0,
+        "test_set": "MLS es_419",
+        "benched": false
+      },
+      "recommended": false
+    },
+    {
+      "path": "es/1120ms",
+      "language": "Spanish",
+      "language_code": "es-ES",
       "chunk_ms": 1120,
+      "latency_s": 1.12,
       "chunk_mel_frames": 112,
-      "pre_encode_cache": 9,
       "total_mel_frames": 121,
-      "att_context_size": [
-        56,
-        0
+      "att_context": [
+        42,
+        13
       ],
-      "cache_channel_shape": [
-        1,
-        24,
-        56,
-        1024
-      ],
-      "cache_time_shape": [
-        1,
-        24,
-        1024,
-        8
-      ],
-      "components": {
-        "preprocessor": {
-          "mlpackage": {
-            "path": "1120ms/preprocessor.mlpackage",
-            "size_bytes": 608879
-          },
-          "mlmodelc": {
-            "path": "1120ms/preprocessor.mlmodelc",
-            "size_bytes": 617594
-          }
-        },
-        "encoder": {
-          "mlpackage": {
-            "path": "1120ms/encoder.mlpackage",
-            "size_bytes": 593966620
-          },
-          "mlmodelc": {
-            "path": "1120ms/encoder.mlmodelc",
-            "size_bytes": 594160802
-          }
-        },
-        "decoder": {
-          "mlpackage": {
-            "path": "1120ms/decoder.mlpackage",
-            "size_bytes": 29881569
-          },
-          "mlmodelc": {
-            "path": "1120ms/decoder.mlmodelc",
-            "size_bytes": 29889159
-          }
-        },
-        "joint": {
-          "mlpackage": {
-            "path": "1120ms/joint.mlpackage",
-            "size_bytes": 18916847
-          },
-          "mlmodelc": {
-            "path": "1120ms/joint.mlmodelc",
-            "size_bytes": 18923548
-          }
-        }
-      },
-      "metadata_path": "1120ms/metadata.json",
-      "tokenizer_path": "1120ms/tokenizer.json",
-      "total_size_bytes": 1287261196
+      "vocab_size": 831,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 106.8,
+        "wer_pct": 6.5,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 2385,
+        "test_set": "MLS es_419",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "es/2240ms",
+      "language": "Spanish",
+      "language_code": "es-ES",
+      "chunk_ms": 2240,
+      "latency_s": 2.24,
+      "chunk_mel_frames": 224,
+      "total_mel_frames": 233,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 831,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 153.6,
+        "wer_pct": 6.8,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 2385,
+        "test_set": "MLS es_419",
+        "benched": true
+      },
+      "recommended": true
+    },
+    {
+      "path": "es/4480ms",
+      "language": "Spanish",
+      "language_code": "es-ES",
+      "chunk_ms": 4480,
+      "latency_s": 4.48,
+      "chunk_mel_frames": 448,
+      "total_mel_frames": 457,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 831,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 139.0,
+        "wer_pct": 6.4,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 2385,
+        "test_set": "MLS es_419",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "fr/560ms",
+      "language": "French",
+      "language_code": "fr-FR",
+      "chunk_ms": 560,
+      "latency_s": 0.56,
+      "chunk_mel_frames": 56,
+      "total_mel_frames": 65,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 848,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": null,
+        "wer_pct": null,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 0,
+        "test_set": "MLS fr_fr",
+        "benched": false
+      },
+      "recommended": false
+    },
+    {
+      "path": "fr/1120ms",
+      "language": "French",
+      "language_code": "fr-FR",
+      "chunk_ms": 1120,
+      "latency_s": 1.12,
+      "chunk_mel_frames": 112,
+      "total_mel_frames": 121,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 848,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 109.5,
+        "wer_pct": 9.9,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 2426,
+        "test_set": "MLS fr_fr",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "fr/2240ms",
+      "language": "French",
+      "language_code": "fr-FR",
+      "chunk_ms": 2240,
+      "latency_s": 2.24,
+      "chunk_mel_frames": 224,
+      "total_mel_frames": 233,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 848,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 134.4,
+        "wer_pct": 10.4,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 2426,
+        "test_set": "MLS fr_fr",
+        "benched": true
+      },
+      "recommended": true
+    },
+    {
+      "path": "fr/4480ms",
+      "language": "French",
+      "language_code": "fr-FR",
+      "chunk_ms": 4480,
+      "latency_s": 4.48,
+      "chunk_mel_frames": 448,
+      "total_mel_frames": 457,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 848,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 130.2,
+        "wer_pct": 16.8,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 100,
+        "test_set": "MLS fr_fr",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "it/560ms",
+      "language": "Italian",
+      "language_code": "it-IT",
+      "chunk_ms": 560,
+      "latency_s": 0.56,
+      "chunk_mel_frames": 56,
+      "total_mel_frames": 65,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 805,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": null,
+        "wer_pct": null,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 0,
+        "test_set": "MLS it_it",
+        "benched": false
+      },
+      "recommended": false
+    },
+    {
+      "path": "it/1120ms",
+      "language": "Italian",
+      "language_code": "it-IT",
+      "chunk_ms": 1120,
+      "latency_s": 1.12,
+      "chunk_mel_frames": 112,
+      "total_mel_frames": 121,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 805,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 109.2,
+        "wer_pct": 23.0,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 1262,
+        "test_set": "MLS it_it",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "it/2240ms",
+      "language": "Italian",
+      "language_code": "it-IT",
+      "chunk_ms": 2240,
+      "latency_s": 2.24,
+      "chunk_mel_frames": 224,
+      "total_mel_frames": 233,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 805,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 136.8,
+        "wer_pct": 18.0,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 1262,
+        "test_set": "MLS it_it",
+        "benched": true
+      },
+      "recommended": true
+    },
+    {
+      "path": "it/4480ms",
+      "language": "Italian",
+      "language_code": "it-IT",
+      "chunk_ms": 4480,
+      "latency_s": 4.48,
+      "chunk_mel_frames": 448,
+      "total_mel_frames": 457,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 805,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 134.6,
+        "wer_pct": 25.6,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 100,
+        "test_set": "MLS it_it",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "pt/560ms",
+      "language": "Portuguese",
+      "language_code": "pt-BR",
+      "chunk_ms": 560,
+      "latency_s": 0.56,
+      "chunk_mel_frames": 56,
+      "total_mel_frames": 65,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 870,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": null,
+        "wer_pct": null,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 0,
+        "test_set": "MLS pt_br",
+        "benched": false
+      },
+      "recommended": false
+    },
+    {
+      "path": "pt/1120ms",
+      "language": "Portuguese",
+      "language_code": "pt-BR",
+      "chunk_ms": 1120,
+      "latency_s": 1.12,
+      "chunk_mel_frames": 112,
+      "total_mel_frames": 121,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 870,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 111.4,
+        "wer_pct": 9.8,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 871,
+        "test_set": "MLS pt_br",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "pt/2240ms",
+      "language": "Portuguese",
+      "language_code": "pt-BR",
+      "chunk_ms": 2240,
+      "latency_s": 2.24,
+      "chunk_mel_frames": 224,
+      "total_mel_frames": 233,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 870,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 155.2,
+        "wer_pct": 12.9,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 871,
+        "test_set": "MLS pt_br",
+        "benched": true
+      },
+      "recommended": true
+    },
+    {
+      "path": "pt/4480ms",
+      "language": "Portuguese",
+      "language_code": "pt-BR",
+      "chunk_ms": 4480,
+      "latency_s": 4.48,
+      "chunk_mel_frames": 448,
+      "total_mel_frames": 457,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 870,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 134.9,
+        "wer_pct": 10.3,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 871,
+        "test_set": "MLS pt_br",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "de/560ms",
+      "language": "German",
+      "language_code": "de-DE",
+      "chunk_ms": 560,
+      "latency_s": 0.56,
+      "chunk_mel_frames": 56,
+      "total_mel_frames": 65,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 795,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": null,
+        "wer_pct": null,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 0,
+        "test_set": "FLEURS de_de",
+        "benched": false
+      },
+      "recommended": false
+    },
+    {
+      "path": "de/1120ms",
+      "language": "German",
+      "language_code": "de-DE",
+      "chunk_ms": 1120,
+      "latency_s": 1.12,
+      "chunk_mel_frames": 112,
+      "total_mel_frames": 121,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 795,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 107.5,
+        "wer_pct": 11.2,
+        "cer_pct": 3.6,
+        "metric": "WER",
+        "n": 862,
+        "test_set": "FLEURS de_de",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "de/2240ms",
+      "language": "German",
+      "language_code": "de-DE",
+      "chunk_ms": 2240,
+      "latency_s": 2.24,
+      "chunk_mel_frames": 224,
+      "total_mel_frames": 233,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 795,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 150.1,
+        "wer_pct": 11.0,
+        "cer_pct": 3.8,
+        "metric": "WER",
+        "n": 862,
+        "test_set": "FLEURS de_de",
+        "benched": true
+      },
+      "recommended": true
+    },
+    {
+      "path": "de/4480ms",
+      "language": "German",
+      "language_code": "de-DE",
+      "chunk_ms": 4480,
+      "latency_s": 4.48,
+      "chunk_mel_frames": 448,
+      "total_mel_frames": 457,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 795,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 151.9,
+        "wer_pct": 11.0,
+        "cer_pct": 3.9,
+        "metric": "WER",
+        "n": 862,
+        "test_set": "FLEURS de_de",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "zh/560ms",
+      "language": "Chinese",
+      "language_code": "zh-CN",
+      "chunk_ms": 560,
+      "latency_s": 0.56,
+      "chunk_mel_frames": 56,
+      "total_mel_frames": 65,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 1875,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": null,
+        "wer_pct": null,
+        "cer_pct": null,
+        "metric": "CER",
+        "n": 0,
+        "test_set": "FLEURS cmn_hans_cn",
+        "benched": false
+      },
+      "recommended": false
+    },
+    {
+      "path": "zh/1120ms",
+      "language": "Chinese",
+      "language_code": "zh-CN",
+      "chunk_ms": 1120,
+      "latency_s": 1.12,
+      "chunk_mel_frames": 112,
+      "total_mel_frames": 121,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 1875,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 106.3,
+        "wer_pct": null,
+        "cer_pct": 21.9,
+        "metric": "CER",
+        "n": 945,
+        "test_set": "FLEURS cmn_hans_cn",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "zh/2240ms",
+      "language": "Chinese",
+      "language_code": "zh-CN",
+      "chunk_ms": 2240,
+      "latency_s": 2.24,
+      "chunk_mel_frames": 224,
+      "total_mel_frames": 233,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 1875,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 146.0,
+        "wer_pct": null,
+        "cer_pct": 21.4,
+        "metric": "CER",
+        "n": 945,
+        "test_set": "FLEURS cmn_hans_cn",
+        "benched": true
+      },
+      "recommended": true
+    },
+    {
+      "path": "zh/4480ms",
+      "language": "Chinese",
+      "language_code": "zh-CN",
+      "chunk_ms": 4480,
+      "latency_s": 4.48,
+      "chunk_mel_frames": 448,
+      "total_mel_frames": 457,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 1875,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 140.2,
+        "wer_pct": null,
+        "cer_pct": 21.4,
+        "metric": "CER",
+        "n": 945,
+        "test_set": "FLEURS cmn_hans_cn",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "ja/560ms",
+      "language": "Japanese",
+      "language_code": "ja-JP",
+      "chunk_ms": 560,
+      "latency_s": 0.56,
+      "chunk_mel_frames": 56,
+      "total_mel_frames": 65,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 1403,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": null,
+        "wer_pct": null,
+        "cer_pct": null,
+        "metric": "CER",
+        "n": 0,
+        "test_set": "FLEURS ja_jp",
+        "benched": false
+      },
+      "recommended": false
+    },
+    {
+      "path": "ja/1120ms",
+      "language": "Japanese",
+      "language_code": "ja-JP",
+      "chunk_ms": 1120,
+      "latency_s": 1.12,
+      "chunk_mel_frames": 112,
+      "total_mel_frames": 121,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 1403,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 108.5,
+        "wer_pct": null,
+        "cer_pct": 15.6,
+        "metric": "CER",
+        "n": 650,
+        "test_set": "FLEURS ja_jp",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "ja/2240ms",
+      "language": "Japanese",
+      "language_code": "ja-JP",
+      "chunk_ms": 2240,
+      "latency_s": 2.24,
+      "chunk_mel_frames": 224,
+      "total_mel_frames": 233,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 1403,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 150.5,
+        "wer_pct": null,
+        "cer_pct": 15.4,
+        "metric": "CER",
+        "n": 650,
+        "test_set": "FLEURS ja_jp",
+        "benched": true
+      },
+      "recommended": true
+    },
+    {
+      "path": "ja/4480ms",
+      "language": "Japanese",
+      "language_code": "ja-JP",
+      "chunk_ms": 4480,
+      "latency_s": 4.48,
+      "chunk_mel_frames": 448,
+      "total_mel_frames": 457,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 1403,
+      "vocab_pruned": true,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "encoder",
+        "joint",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 147.8,
+        "wer_pct": null,
+        "cer_pct": 15.3,
+        "metric": "CER",
+        "n": 650,
+        "test_set": "FLEURS ja_jp",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "multilingual/560ms",
+      "language": "Multilingual (100+ langs via prompt_id)",
+      "language_code": "auto",
+      "chunk_ms": 560,
+      "latency_s": 0.56,
+      "chunk_mel_frames": 56,
+      "total_mel_frames": 65,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 13087,
+      "vocab_pruned": false,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": null,
+        "wer_pct": null,
+        "cer_pct": null,
+        "metric": "WER",
+        "n": 0,
+        "test_set": "LibriSpeech test-clean (en)",
+        "benched": false
+      },
+      "recommended": false
+    },
+    {
+      "path": "multilingual/1120ms",
+      "language": "Multilingual (100+ langs via prompt_id)",
+      "language_code": "auto",
+      "chunk_ms": 1120,
+      "latency_s": 1.12,
+      "chunk_mel_frames": 112,
+      "total_mel_frames": 121,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 13087,
+      "vocab_pruned": false,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 62.2,
+        "wer_pct": 3.8,
+        "cer_pct": 1.4,
+        "metric": "WER",
+        "n": 2620,
+        "test_set": "LibriSpeech test-clean (en)",
+        "benched": true
+      },
+      "recommended": false
+    },
+    {
+      "path": "multilingual/2240ms",
+      "language": "Multilingual (100+ langs via prompt_id)",
+      "language_code": "auto",
+      "chunk_ms": 2240,
+      "latency_s": 2.24,
+      "chunk_mel_frames": 224,
+      "total_mel_frames": 233,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 13087,
+      "vocab_pruned": false,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 74.6,
+        "wer_pct": 3.7,
+        "cer_pct": 1.4,
+        "metric": "WER",
+        "n": 2620,
+        "test_set": "LibriSpeech test-clean (en)",
+        "benched": true
+      },
+      "recommended": true
+    },
+    {
+      "path": "multilingual/4480ms",
+      "language": "Multilingual (100+ langs via prompt_id)",
+      "language_code": "auto",
+      "chunk_ms": 4480,
+      "latency_s": 4.48,
+      "chunk_mel_frames": 448,
+      "total_mel_frames": 457,
+      "att_context": [
+        42,
+        13
+      ],
+      "vocab_size": 13087,
+      "vocab_pruned": false,
+      "components": [
+        "decoder",
+        "decoder_joint",
+        "decoder_joint_noencproj",
+        "encoder",
+        "joint",
+        "joint_noencproj_batched",
+        "preprocessor"
+      ],
+      "formats": [
+        "mlpackage",
+        "mlmodelc"
+      ],
+      "benchmark": {
+        "rtfx": 19.4,
+        "wer_pct": 3.7,
+        "cer_pct": 1.4,
+        "metric": "WER",
+        "n": 2620,
+        "test_set": "LibriSpeech test-clean (en)",
+        "benched": true
+      },
+      "recommended": false
     }
   ]
-}
+}

multilingual/1120ms/decoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c32520b84ded2c698000854a77228adf394db522b5a3c25f7737415aae7ed0d
+size 243

multilingual/1120ms/decoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa812bb65dd2a3bef6acf584b2abd5d0f26f4d09afaccf6c5dfd41e630d0fd1b
+size 433

multilingual/1120ms/decoder.mlmodelc/model.mil

@@ -0,0 +1,64 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)];
+            int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)];
+            bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [13088, 640]> module_prediction_embed_weight_to_fp16 = const()[name = string("module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [13088, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_8")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)];
+            int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)];
+            string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")];
+            int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)];
+            int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)];
+            string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            string input_lstm_layer_0_direction_0 = const()[name = string("input_lstm_layer_0_direction_0"), val = string("forward")];
+            bool input_lstm_layer_0_output_sequence_0 = const()[name = string("input_lstm_layer_0_output_sequence_0"), val = bool(true)];
+            string input_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")];
+            string input_lstm_layer_0_cell_activation_0 = const()[name = string("input_lstm_layer_0_cell_activation_0"), val = string("tanh")];
+            string input_lstm_layer_0_activation_0 = const()[name = string("input_lstm_layer_0_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16752768)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20029632)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23306496)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 640]> input_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_lstm_layer_0_cast_fp16_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_lstm_c0_squeeze_cast_fp16")];
+            string input_direction_0 = const()[name = string("input_direction_0"), val = string("forward")];
+            bool input_output_sequence_0 = const()[name = string("input_output_sequence_0"), val = bool(true)];
+            string input_recurrent_activation_0 = const()[name = string("input_recurrent_activation_0"), val = string("sigmoid")];
+            string input_cell_activation_0 = const()[name = string("input_cell_activation_0"), val = string("tanh")];
+            string input_activation_0 = const()[name = string("input_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23311680)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26588544)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29865408)))];
+            tensor<fp16, [1, 1, 640]> input_cast_fp16_0, tensor<fp16, [1, 640]> input_cast_fp16_1, tensor<fp16, [1, 640]> input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_3_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_lstm_layer_0_cast_fp16_0)[name = string("input_cast_fp16")];
+            int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_cast_fp16_1, input_cast_fp16_1))[name = string("obj_3_cast_fp16")];
+            string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_lstm_layer_0_cast_fp16_2, input_cast_fp16_2))[name = string("obj_cast_fp16")];
+            string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
+            string transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = string("transpose_0_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp16, [1, 640, 1]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = string("transpose_1")];
+            tensor<fp32, [1, 640, 1]> decoder_out = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = string("cast_3")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")];
+        } -> (decoder_out, h_out, c_out);
+}

multilingual/1120ms/decoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e49d029739602d265719f0040fcf5328c007a1ed62d0c6ea621e0b2aaeb9a64
+size 29870592

multilingual/1120ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c26ee345b7763ed9f217561572b1386719956de5b58e9174f5586926b4ab85c5
+size 10360

multilingual/1120ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e49d029739602d265719f0040fcf5328c007a1ed62d0c6ea621e0b2aaeb9a64
+size 29870592

multilingual/1120ms/decoder.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "542DC13B-08DF-47C7-AAAA-C2F9DE67BB37": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "8B23B00A-4F60-49E4-B460-719FB6B05887": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "8B23B00A-4F60-49E4-B460-719FB6B05887"
+}

multilingual/1120ms/decoder_joint.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b314763ced4d2bd27484b8ec2a9c60939b724f8ab60b32d29ad0c03f6192599
+size 243

multilingual/1120ms/decoder_joint.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:729baa5678fde0b9fa3e46044cb8eafcb96249bff4c306740e1c40ce326b7101
+size 454

multilingual/1120ms/decoder_joint.mlmodelc/model.mil

@@ -0,0 +1,83 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [1, 1024, 1]> encoder, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)];
+            int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)];
+            bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [13088, 640]> decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [13088, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_9")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)];
+            int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)];
+            string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_8")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")];
+            int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)];
+            int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)];
+            string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")];
+            bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)];
+            string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")];
+            string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16752768)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20029632)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23306496)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_4")];
+            tensor<fp16, [1, 1, 640]> input_5_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")];
+            string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")];
+            bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)];
+            string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")];
+            string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23311680)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26588544)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29865408)))];
+            tensor<fp16, [1, 1, 640]> input_5_cast_fp16_0, tensor<fp16, [1, 640]> input_5_cast_fp16_1, tensor<fp16, [1, 640]> input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")];
+            int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")];
+            string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")];
+            string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [3]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [3]> input_7_perm_0 = const()[name = string("input_7_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [640, 1024]> joint_module_enc_weight_to_fp16 = const()[name = string("joint_module_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29870592)))];
+            tensor<fp16, [640]> joint_module_enc_bias_to_fp16 = const()[name = string("joint_module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31181376)))];
+            tensor<fp16, [1, 1024, 1]> encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_4")];
+            tensor<fp16, [1, 1, 1024]> input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = encoder_to_fp16)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31182720)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32001984)))];
+            tensor<fp16, [1, 1, 640]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_3")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_79_axes_0 = const()[name = string("op_79_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_79_cast_fp16 = expand_dims(axes = var_79_axes_0, x = linear_0_cast_fp16)[name = string("op_79_cast_fp16")];
+            tensor<int32, [1]> var_80_axes_0 = const()[name = string("op_80_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_80_cast_fp16 = expand_dims(axes = var_80_axes_0, x = linear_1_cast_fp16)[name = string("op_80_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_11_cast_fp16 = add(x = var_79_cast_fp16, y = var_80_cast_fp16)[name = string("input_11_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_13_cast_fp16 = relu(x = input_11_cast_fp16)[name = string("input_13_cast_fp16")];
+            tensor<fp16, [13088, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [13088, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32003328)))];
+            tensor<fp16, [13088]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [13088]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48756032)))];
+            tensor<fp16, [1, 1, 1, 13088]> linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_2_cast_fp16")];
+            string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 1, 1, 13088]> logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_3")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_5")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_6")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")];
+        } -> (logits, h_out, c_out);
+}

multilingual/1120ms/decoder_joint.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f283edec035d616e9e1372419dd9bfd8a2de2c92a70b23f6a0f5ed93366ebb03
+size 48782272

multilingual/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:507c3a291a78a11f62b898c64e611016f518d1af658b2aa55c054e0a1029f7ea
+size 13746

multilingual/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f283edec035d616e9e1372419dd9bfd8a2de2c92a70b23f6a0f5ed93366ebb03
+size 48782272

multilingual/1120ms/decoder_joint.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "627E3113-852A-47BD-981E-FAB26C6AB6D0": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "7EEB6F52-3184-47E3-98CD-28268604F7F1": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "7EEB6F52-3184-47E3-98CD-28268604F7F1"
+}

multilingual/1120ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4e3c25c06d72ba514e93ae2ea0dd313f057622a3fb70f65de3bee4b80b3946b
+size 243

multilingual/1120ms/decoder_joint_noencproj.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c9ea28113904cfeca8e5684fb8b54358eb48338f26c894739e3bd076848dcadd
+size 519

multilingual/1120ms/decoder_joint_noencproj.mlmodelc/model.mil

@@ -0,0 +1,91 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios18>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [1, 1, 640]> encoder_proj, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)];
+            bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [13088, 640]> decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [13088, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")];
+            string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_10")];
+            tensor<int32, [1, 1]> cast_1 = cast(dtype = cast_1_dtype_0, x = token_to_int16)[name = string("cast_9")];
+            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
+            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(13088)];
+            tensor<int32, [1, 1]> add_2 = add(x = cast_1, y = slice_by_index_0)[name = string("add_2")];
+            tensor<int32, [1, 1]> select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = string("select_0")];
+            int32 y_cast_fp16_cast_uint16_axis_0 = const()[name = string("y_cast_fp16_cast_uint16_axis_0"), val = int32(0)];
+            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
+            tensor<int16, [1, 1]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_8")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)];
+            int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)];
+            string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")];
+            int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)];
+            int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)];
+            string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")];
+            bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)];
+            string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")];
+            string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16752768)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20029632)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23306496)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_3")];
+            tensor<fp16, [1, 1, 640]> input_5_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")];
+            string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")];
+            bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)];
+            string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")];
+            string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23311680)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26588544)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29865408)))];
+            tensor<fp16, [1, 1, 640]> input_5_cast_fp16_0, tensor<fp16, [1, 640]> input_5_cast_fp16_1, tensor<fp16, [1, 640]> input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")];
+            int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")];
+            string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")];
+            string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [3]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29870592)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30689856)))];
+            tensor<fp16, [1, 1, 640]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [1]> f_axes_0 = const()[name = string("f_axes_0"), val = tensor<int32, [1]>([2])];
+            string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 1, 640]> encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_3")];
+            tensor<fp16, [1, 1, 1, 640]> f_cast_fp16 = expand_dims(axes = f_axes_0, x = encoder_proj_to_fp16)[name = string("f_cast_fp16")];
+            tensor<int32, [1]> g_axes_0 = const()[name = string("g_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> g_cast_fp16 = expand_dims(axes = g_axes_0, x = linear_0_cast_fp16)[name = string("g_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_9_cast_fp16 = add(x = f_cast_fp16, y = g_cast_fp16)[name = string("input_9_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_11_cast_fp16 = relu(x = input_9_cast_fp16)[name = string("input_11_cast_fp16")];
+            tensor<fp16, [13088, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [13088, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30691200)))];
+            tensor<fp16, [13088]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [13088]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47443904)))];
+            tensor<fp16, [1, 1, 1, 13088]> linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_1_cast_fp16")];
+            int32 var_83 = const()[name = string("op_83"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1, 13088]> var_85_softmax_cast_fp16 = softmax(axis = var_83, x = linear_1_cast_fp16)[name = string("op_85_softmax_cast_fp16")];
+            fp32 var_85_epsilon_0 = const()[name = string("op_85_epsilon_0"), val = fp32(0x1p-149)];
+            tensor<fp16, [1, 1, 1, 13088]> var_85_cast_fp16 = log(epsilon = var_85_epsilon_0, x = var_85_softmax_cast_fp16)[name = string("op_85_cast_fp16")];
+            string var_85_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_85_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 1, 1, 13088]> logits = cast(dtype = var_85_cast_fp16_to_fp32_dtype_0, x = var_85_cast_fp16)[name = string("cast_2")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")];
+        } -> (logits, h_out, c_out);
+}

multilingual/1120ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c7ddfe0cb3be2e2896258d91a95483b898e8a274c49fee256e8effd86dc64dda
+size 47470144

multilingual/1120ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0cd7f3fdcb96eec2c46b11da7cb758c2c7ec851677cc1d089bf0388138761c22
+size 14631

multilingual/1120ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c7ddfe0cb3be2e2896258d91a95483b898e8a274c49fee256e8effd86dc64dda
+size 47470144

multilingual/1120ms/decoder_joint_noencproj.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "0D1E34D7-EC88-4977-9F50-17C13013772E": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "4F1B4339-60A2-4F3E-9F0C-B5B2CB77BDAE": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "4F1B4339-60A2-4F3E-9F0C-B5B2CB77BDAE"
+}

multilingual/1120ms/encoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:afb78b1b16fbeee7e0c4953a6401b77ed7b9fa6e27ba9c55ac386ca27e1033cf
+size 243

multilingual/1120ms/encoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e316a70a3a1d1d51edcee4b8c6f386d40684c8892cf84c5d2c478c10458ce6d
+size 572

multilingual/1120ms/encoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f23e8a17303bd5838e6ad4a402d26552f06a3097834d9fc4bf8e11973e63aa7
+size 564607680

multilingual/1120ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66717fbcd7c596e73f7da91a275ba5d7f9f808ed141cc3f565f03f0565882926
+size 800661

multilingual/1120ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f23e8a17303bd5838e6ad4a402d26552f06a3097834d9fc4bf8e11973e63aa7
+size 564607680

multilingual/1120ms/encoder.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "1ED77C41-E152-4A99-A249-1381B7A2B0B2": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "3DFEC7C7-7F84-47C7-AA9B-03492E557363": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "3DFEC7C7-7F84-47C7-AA9B-03492E557363"
+}

multilingual/1120ms/joint.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a285566495ca331c28bd65cb8cd6869402e41b7e1acaca31afd91458a2070130
+size 243

multilingual/1120ms/joint.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:38ad7913a24486cc3178df1a42d6a8233bcd54c5b42f59bc419ff9101bd19135
+size 341

multilingual/1120ms/joint.mlmodelc/model.mil

@@ -0,0 +1,31 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [1, 640, 1]> decoder, tensor<fp32, [1, 1024, 1]> encoder) {
+            tensor<int32, [3]> input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [640, 1024]> module_enc_weight_to_fp16 = const()[name = string("module_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [640]> module_enc_bias_to_fp16 = const()[name = string("module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1310848)))];
+            tensor<fp16, [1, 1024, 1]> encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_2")];
+            tensor<fp16, [1, 1, 1024]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_to_fp16)[name = string("transpose_1")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = module_enc_bias_to_fp16, weight = module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> module_pred_weight_to_fp16 = const()[name = string("module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1312192)))];
+            tensor<fp16, [640]> module_pred_bias_to_fp16 = const()[name = string("module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2131456)))];
+            tensor<fp16, [1, 640, 1]> decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_1")];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_to_fp16)[name = string("transpose_0")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = module_pred_bias_to_fp16, weight = module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_23_axes_0 = const()[name = string("op_23_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = string("op_23_cast_fp16")];
+            tensor<int32, [1]> var_25_axes_0 = const()[name = string("op_25_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_25_cast_fp16 = expand_dims(axes = var_25_axes_0, x = linear_1_cast_fp16)[name = string("op_25_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_25_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("input_7_cast_fp16")];
+            tensor<fp16, [13088, 640]> module_joint_net_2_weight_to_fp16 = const()[name = string("module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [13088, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2132800)))];
+            tensor<fp16, [13088]> module_joint_net_2_bias_to_fp16 = const()[name = string("module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [13088]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18885504)))];
+            tensor<fp16, [1, 1, 1, 13088]> linear_2_cast_fp16 = linear(bias = module_joint_net_2_bias_to_fp16, weight = module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_2_cast_fp16")];
+            string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 1, 1, 13088]> logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_0")];
+        } -> (logits);
+}

multilingual/1120ms/joint.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4c8ae93e304a187ebfa0b88c812b70e79b625a549727922e7f63d61c1c7b6dd
+size 18911744

multilingual/1120ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8897a4790a6c88fde37843c50b4c45d5ea24f24ea5811fe7555b54fcdde8a5c0
+size 4486

multilingual/1120ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4c8ae93e304a187ebfa0b88c812b70e79b625a549727922e7f63d61c1c7b6dd
+size 18911744

multilingual/1120ms/joint.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "3F33ED2D-8E5C-4BBD-B243-7D003571E7E2": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "7D35F675-3334-491B-8264-00E768D11202": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "3F33ED2D-8E5C-4BBD-B243-7D003571E7E2"
+}

multilingual/1120ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5f4a9b0771be9af64fd93db7ceb42dbd305920b1260fe2219f1b046e84841cd
+size 243

multilingual/1120ms/joint_noencproj_batched.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc07c4c2de2b13127f406ee70373b2c178702a03755bdc7a2bd57e623b5e65c5
+size 406

multilingual/1120ms/joint_noencproj_batched.mlmodelc/model.mil

@@ -0,0 +1,26 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios18>(tensor<fp32, [1, 640, 1]> decoder, tensor<fp32, [1, 4, 640]> encoder_proj) {
+            tensor<int32, [3]> input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819328)))];
+            tensor<fp16, [1, 640, 1]> decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_2")];
+            tensor<fp16, [1, 1, 640]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = decoder_to_fp16)[name = string("transpose_0")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [1]> var_15_axes_0 = const()[name = string("op_15_axes_0"), val = tensor<int32, [1]>([2])];
+            string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 4, 640]> encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_1")];
+            tensor<fp16, [1, 4, 1, 640]> var_15_cast_fp16 = expand_dims(axes = var_15_axes_0, x = encoder_proj_to_fp16)[name = string("op_15_cast_fp16")];
+            tensor<int32, [1]> var_17_axes_0 = const()[name = string("op_17_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_17_cast_fp16 = expand_dims(axes = var_17_axes_0, x = linear_0_cast_fp16)[name = string("op_17_cast_fp16")];
+            tensor<fp16, [1, 4, 1, 640]> input_3_cast_fp16 = add(x = var_15_cast_fp16, y = var_17_cast_fp16)[name = string("input_3_cast_fp16")];
+            tensor<fp16, [1, 4, 1, 640]> input_5_cast_fp16 = relu(x = input_3_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<fp16, [13088, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [13088, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(820672)))];
+            tensor<fp16, [13088]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [13088]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17573376)))];
+            tensor<fp16, [1, 4, 1, 13088]> linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_5_cast_fp16)[name = string("linear_1_cast_fp16")];
+            string linear_1_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_1_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 4, 1, 13088]> logits = cast(dtype = linear_1_cast_fp16_to_fp32_dtype_0, x = linear_1_cast_fp16)[name = string("cast_0")];
+        } -> (logits);
+}

multilingual/1120ms/joint_noencproj_batched.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d6b104e9d6990c07d6cd41bafe27cae8d39cfe037ec701584c47af1094daeeb
+size 17599616

multilingual/1120ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ebb2680467cb6774940b609a57ff38df74262ab683052ee3615204e4bd1e3949
+size 3840

multilingual/1120ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d6b104e9d6990c07d6cd41bafe27cae8d39cfe037ec701584c47af1094daeeb
+size 17599616

multilingual/1120ms/joint_noencproj_batched.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "55B094CA-55E5-480A-8B14-30A24DC3EEF0": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "CA02FD13-87CE-4425-9B49-DE8265EC1B54": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "CA02FD13-87CE-4425-9B49-DE8265EC1B54"
+}

multilingual/1120ms/metadata.json

@@ -0,0 +1,196 @@
+{
+  "model": "nvidia/nemotron-asr-streaming-multilingual-0.6b",
+  "model_class": "nemo.collections.asr.models.rnnt_bpe_models_prompt.EncDecRNNTBPEModelWithPrompt",
+  "sample_rate": 16000,
+  "mel_features": 128,
+  "chunk_mel_frames": 112,
+  "pre_encode_cache": 9,
+  "total_mel_frames": 121,
+  "att_context_size": [
+    42,
+    13
+  ],
+  "vocab_size": 13087,
+  "blank_idx": 13087,
+  "cache_channel_shape": [
+    1,
+    24,
+    42,
+    1024
+  ],
+  "cache_time_shape": [
+    1,
+    24,
+    1024,
+    8
+  ],
+  "decoder_hidden": 640,
+  "decoder_layers": 2,
+  "encoder_dim": 1024,
+  "num_prompts": 128,
+  "prompt_dictionary": {
+    "af-ZA": 54,
+    "am-ET": 49,
+    "ar": 7,
+    "ar-AR": 7,
+    "auto": 101,
+    "ay-BO": 81,
+    "az-AZ": 66,
+    "bg": 30,
+    "bg-BG": 30,
+    "bn-IN": 36,
+    "cs": 22,
+    "cs-CZ": 22,
+    "da": 25,
+    "da-DK": 25,
+    "de": 9,
+    "de-DE": 9,
+    "el": 21,
+    "el-GR": 21,
+    "en": 0,
+    "en-GB": 1,
+    "en-US": 0,
+    "enGB": 1,
+    "es": 3,
+    "es-ES": 2,
+    "es-US": 3,
+    "esES": 2,
+    "et": 60,
+    "et-EE": 60,
+    "fa-IR": 38,
+    "fi": 26,
+    "fi-FI": 26,
+    "fr": 8,
+    "fr-CA": 100,
+    "fr-FR": 8,
+    "gn-PY": 82,
+    "gu-IN": 42,
+    "ha-NG": 50,
+    "haw-US": 97,
+    "he-IL": 64,
+    "hi": 6,
+    "hi-HI": 6,
+    "hi-IN": 6,
+    "hr": 29,
+    "hr-HR": 29,
+    "hu": 23,
+    "hu-HU": 23,
+    "hy-AM": 68,
+    "id-ID": 34,
+    "ig-NG": 53,
+    "it": 15,
+    "it-IT": 15,
+    "ja-JA": 10,
+    "ja-JP": 10,
+    "ka-GE": 67,
+    "km-KH": 47,
+    "kn-IN": 43,
+    "ko": 14,
+    "ko-KO": 14,
+    "ko-KR": 14,
+    "ku-TR": 65,
+    "ky-KG": 71,
+    "ln-CD": 58,
+    "lt": 31,
+    "lt-LT": 31,
+    "lv": 61,
+    "lv-LV": 61,
+    "mi-NZ": 96,
+    "ml-IN": 44,
+    "mr-IN": 41,
+    "ms-MY": 35,
+    "mt-MT": 102,
+    "nah-MX": 83,
+    "nb": 103,
+    "nb-NO": 103,
+    "ne-NP": 46,
+    "nl": 16,
+    "nl-NL": 16,
+    "nn": 104,
+    "nn-NO": 104,
+    "no": 27,
+    "no-NO": 27,
+    "ny-MW": 57,
+    "or-KE": 59,
+    "pl": 17,
+    "pl-PL": 17,
+    "pt": 13,
+    "pt-BR": 12,
+    "pt-PT": 13,
+    "qu-PE": 80,
+    "ro": 20,
+    "ro-RO": 20,
+    "ru": 11,
+    "ru-RU": 11,
+    "rw-RW": 55,
+    "si-LK": 45,
+    "sk": 28,
+    "sk-SK": 28,
+    "sl": 62,
+    "sl-SI": 62,
+    "sm-WS": 98,
+    "so-SO": 56,
+    "sv": 24,
+    "sv-SE": 24,
+    "sw-KE": 48,
+    "ta-IN": 39,
+    "te-IN": 40,
+    "tg-TJ": 70,
+    "th-TH": 32,
+    "to-TO": 99,
+    "tr": 18,
+    "tr-TR": 18,
+    "uk": 19,
+    "uk-UA": 19,
+    "ur-PK": 37,
+    "uz-UZ": 69,
+    "vi-VN": 33,
+    "yo-NG": 52,
+    "zh-CN": 4,
+    "zh-TW": 5,
+    "zh-ZH": 4,
+    "zu-ZA": 51
+  },
+  "default_prompt_id": 101,
+  "lang_tag_token_ids": [
+    1,
+    256,
+    397,
+    518,
+    673,
+    814,
+    907,
+    993,
+    1125,
+    1232,
+    1279,
+    1383,
+    1455,
+    1603,
+    1724,
+    1841,
+    1929,
+    2021,
+    2124,
+    2205,
+    2322,
+    2440,
+    2529,
+    2809,
+    2947,
+    2986,
+    3051,
+    3064,
+    3134,
+    3247,
+    3446,
+    7489,
+    9532,
+    9544,
+    9596,
+    9695,
+    9815,
+    9847,
+    12944
+  ]
+}

multilingual/1120ms/preprocessor.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:63c74dcaace5d0cef0b6bcd65225e8985e605517fa97f95b13218d02735b6a42
+size 243

multilingual/1120ms/preprocessor.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7b11e08aba46d1845d8ad3f247717e0f6fae35b21d71d52e44a69ea73587bfe
+size 371

multilingual/1120ms/preprocessor.mlmodelc/model.mil

@@ -0,0 +1,122 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [1, ?]> audio, tensor<int32, [1]> audio_length) [FlexibleShapeInformation = tuple<tuple<string, dict<string, tensor<int32, [?]>>>, tuple<string, dict<string, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"audio", [1, 1]}}), ("RangeDims", {{"audio", [[1, 1], [1, 480000]]}})))] {
+            int32 var_9 = const()[name = string("op_9"), val = int32(1)];
+            int32 var_10 = const()[name = string("op_10"), val = int32(160)];
+            int32 var_12 = const()[name = string("op_12"), val = int32(0)];
+            int32 var_33 = const()[name = string("op_33"), val = int32(512)];
+            tensor<int32, [1]> var_34 = add(x = audio_length, y = var_33)[name = string("op_34")];
+            int32 var_35 = const()[name = string("op_35"), val = int32(512)];
+            tensor<int32, [1]> var_36 = sub(x = var_34, y = var_35)[name = string("op_36")];
+            tensor<int32, [1]> floor_div_0 = floor_div(x = var_36, y = var_10)[name = string("floor_div_0")];
+            tensor<bool, [1]> var_39 = equal(x = audio_length, y = var_12)[name = string("op_39")];
+            tensor<int32, [1]> var_40 = const()[name = string("op_40"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> mel_length = select(a = var_40, b = floor_div_0, cond = var_39)[name = string("seq_len")];
+            string audio_to_fp16_dtype_0 = const()[name = string("audio_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, ?]> audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = string("cast_14")];
+            tensor<int32, [2]> var_42_shape_cast_fp16 = shape(x = audio_to_fp16)[name = string("op_42_shape_cast_fp16")];
+            int32 gather_0_axis_0 = const()[name = string("gather_0_axis_0"), val = int32(0)];
+            int32 gather_0_batch_dims_0 = const()[name = string("gather_0_batch_dims_0"), val = int32(0)];
+            bool gather_0_validate_indices_0 = const()[name = string("gather_0_validate_indices_0"), val = bool(false)];
+            string var_42_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_42_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")];
+            uint16 select_0_to_uint16 = const()[name = string("select_0_to_uint16"), val = uint16(1)];
+            tensor<int16, [2]> var_42_shape_cast_fp16_to_int16 = cast(dtype = var_42_shape_cast_fp16_to_int16_dtype_0, x = var_42_shape_cast_fp16)[name = string("cast_13")];
+            int16 gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_42_shape_cast_fp16_to_int16)[name = string("gather_0_cast_uint16")];
+            string gather_0_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_0_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            int32 const_0 = const()[name = string("const_0"), val = int32(0)];
+            int32 const_1 = const()[name = string("const_1"), val = int32(1)];
+            int32 gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = string("cast_12")];
+            tensor<int32, [?]> var_43 = range_1d(end = gather_0_cast_uint16_to_int32, start = const_0, step = const_1)[name = string("op_43")];
+            tensor<int32, [1]> var_44_axes_0 = const()[name = string("op_44_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = string("op_44")];
+            tensor<int32, [1]> var_45_axes_0 = const()[name = string("op_45_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_45 = expand_dims(axes = var_45_axes_0, x = audio_length)[name = string("op_45")];
+            tensor<bool, [1, ?]> timemask = less(x = var_44, y = var_45)[name = string("timemask")];
+            tensor<int32, [2]> var_48_begin_0 = const()[name = string("op_48_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_48_end_0 = const()[name = string("op_48_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> var_48_end_mask_0 = const()[name = string("op_48_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<bool, [2]> var_48_squeeze_mask_0 = const()[name = string("op_48_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
+            tensor<fp16, [1]> var_48_cast_fp16 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, squeeze_mask = var_48_squeeze_mask_0, x = audio_to_fp16)[name = string("op_48_cast_fp16")];
+            tensor<int32, [1]> var_49_axes_0 = const()[name = string("op_49_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1]> var_49_cast_fp16 = expand_dims(axes = var_49_axes_0, x = var_48_cast_fp16)[name = string("op_49_cast_fp16")];
+            tensor<int32, [2]> var_51_begin_0 = const()[name = string("op_51_begin_0"), val = tensor<int32, [2]>([0, 1])];
+            tensor<int32, [2]> var_51_end_0 = const()[name = string("op_51_end_0"), val = tensor<int32, [2]>([1, 0])];
+            tensor<bool, [2]> var_51_end_mask_0 = const()[name = string("op_51_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<fp16, [1, ?]> var_51_cast_fp16 = slice_by_index(begin = var_51_begin_0, end = var_51_end_0, end_mask = var_51_end_mask_0, x = audio_to_fp16)[name = string("op_51_cast_fp16")];
+            tensor<int32, [2]> var_53_begin_0 = const()[name = string("op_53_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_53_end_0 = const()[name = string("op_53_end_0"), val = tensor<int32, [2]>([1, -1])];
+            tensor<bool, [2]> var_53_end_mask_0 = const()[name = string("op_53_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp16, [1, ?]> var_53_cast_fp16 = slice_by_index(begin = var_53_begin_0, end = var_53_end_0, end_mask = var_53_end_mask_0, x = audio_to_fp16)[name = string("op_53_cast_fp16")];
+            fp16 var_54_to_fp16 = const()[name = string("op_54_to_fp16"), val = fp16(0x1.f0cp-1)];
+            tensor<fp16, [1, ?]> var_55_cast_fp16 = mul(x = var_53_cast_fp16, y = var_54_to_fp16)[name = string("op_55_cast_fp16")];
+            tensor<fp16, [1, ?]> var_56_cast_fp16 = sub(x = var_51_cast_fp16, y = var_55_cast_fp16)[name = string("op_56_cast_fp16")];
+            bool x_3_interleave_0 = const()[name = string("x_3_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, ?]> x_3_cast_fp16 = concat(axis = var_9, interleave = x_3_interleave_0, values = (var_49_cast_fp16, var_56_cast_fp16))[name = string("x_3_cast_fp16")];
+            tensor<bool, [1, ?]> var_59 = logical_not(x = timemask)[name = string("op_59")];
+            fp16 var_16_to_fp16 = const()[name = string("op_16_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, ?]> input_1_cast_fp16 = select(a = var_16_to_fp16, b = x_3_cast_fp16, cond = var_59)[name = string("input_1_cast_fp16")];
+            tensor<int32, [3]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, ?]> input_3_cast_fp16 = reshape(shape = concat_1x, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")];
+            tensor<int32, [6]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 256, 256])];
+            string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("constant")];
+            fp16 const_3_to_fp16 = const()[name = string("const_3_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, ?]> input_5_cast_fp16 = pad(constant_val = const_3_to_fp16, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<int32, [2]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [2]>([1, -1])];
+            tensor<fp16, [1, ?]> input_cast_fp16 = reshape(shape = concat_2x, x = input_5_cast_fp16)[name = string("input_cast_fp16")];
+            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([160])];
+            tensor<int32, [1]> expand_dims_4_axes_0 = const()[name = string("expand_dims_4_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, ?]> expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = input_cast_fp16)[name = string("expand_dims_4_cast_fp16")];
+            string conv_0_pad_type_0 = const()[name = string("conv_0_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> conv_0_pad_0 = const()[name = string("conv_0_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_0_dilations_0 = const()[name = string("conv_0_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 conv_0_groups_0 = const()[name = string("conv_0_groups_0"), val = int32(1)];
+            tensor<fp16, [257, 1, 512]> expand_dims_1_to_fp16 = const()[name = string("expand_dims_1_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 257, ?]> conv_0_cast_fp16 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_0_cast_fp16")];
+            string conv_1_pad_type_0 = const()[name = string("conv_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> conv_1_pad_0 = const()[name = string("conv_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_1_dilations_0 = const()[name = string("conv_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 conv_1_groups_0 = const()[name = string("conv_1_groups_0"), val = int32(1)];
+            tensor<fp16, [257, 1, 512]> expand_dims_2_to_fp16 = const()[name = string("expand_dims_2_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263296)))];
+            tensor<fp16, [1, 257, ?]> conv_1_cast_fp16 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_1_cast_fp16")];
+            int32 stack_0_axis_0 = const()[name = string("stack_0_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 257, ?, 2]> stack_0_cast_fp16 = stack(axis = stack_0_axis_0, values = (conv_0_cast_fp16, conv_1_cast_fp16))[name = string("stack_0_cast_fp16")];
+            fp16 var_19_promoted_to_fp16 = const()[name = string("op_19_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1, 257, ?, 2]> var_74_cast_fp16 = pow(x = stack_0_cast_fp16, y = var_19_promoted_to_fp16)[name = string("op_74_cast_fp16")];
+            tensor<int32, [1]> var_76_axes_0 = const()[name = string("op_76_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_76_keep_dims_0 = const()[name = string("op_76_keep_dims_0"), val = bool(false)];
+            tensor<fp16, [1, 257, ?]> var_76_cast_fp16 = reduce_sum(axes = var_76_axes_0, keep_dims = var_76_keep_dims_0, x = var_74_cast_fp16)[name = string("op_76_cast_fp16")];
+            tensor<fp16, [1, 257, ?]> x_11_cast_fp16 = identity(x = var_76_cast_fp16)[name = string("x_11_cast_fp16")];
+            bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)];
+            bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 128, 257]> const_4_to_fp16 = const()[name = string("const_4_to_fp16"), val = tensor<fp16, [1, 128, 257]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526528)))];
+            tensor<fp16, [1, 128, ?]> x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = const_4_to_fp16, y = x_11_cast_fp16)[name = string("x_13_cast_fp16")];
+            fp16 var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, ?]> var_84_cast_fp16 = add(x = x_13_cast_fp16, y = var_83_to_fp16)[name = string("op_84_cast_fp16")];
+            fp32 x_epsilon_0 = const()[name = string("x_epsilon_0"), val = fp32(0x1p-149)];
+            tensor<fp16, [1, 128, ?]> x_cast_fp16 = log(epsilon = x_epsilon_0, x = var_84_cast_fp16)[name = string("x_cast_fp16")];
+            tensor<int32, [3]> var_86_shape_cast_fp16 = shape(x = x_cast_fp16)[name = string("op_86_shape_cast_fp16")];
+            int32 gather_5_axis_0 = const()[name = string("gather_5_axis_0"), val = int32(0)];
+            int32 gather_5_batch_dims_0 = const()[name = string("gather_5_batch_dims_0"), val = int32(0)];
+            bool gather_5_validate_indices_0 = const()[name = string("gather_5_validate_indices_0"), val = bool(false)];
+            string var_86_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_86_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
+            uint16 select_5_to_uint16 = const()[name = string("select_5_to_uint16"), val = uint16(2)];
+            tensor<uint16, [3]> var_86_shape_cast_fp16_to_uint16 = cast(dtype = var_86_shape_cast_fp16_to_uint16_dtype_0, x = var_86_shape_cast_fp16)[name = string("cast_11")];
+            uint16 gather_5_cast_uint16 = gather(axis = gather_5_axis_0, batch_dims = gather_5_batch_dims_0, indices = select_5_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_86_shape_cast_fp16_to_uint16)[name = string("gather_5_cast_uint16")];
+            string gather_5_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_5_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            int32 const_5 = const()[name = string("const_5"), val = int32(0)];
+            int32 const_6 = const()[name = string("const_6"), val = int32(1)];
+            int32 gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16)[name = string("cast_10")];
+            tensor<int32, [?]> mask_1 = range_1d(end = gather_5_cast_uint16_to_int32, start = const_5, step = const_6)[name = string("mask_1")];
+            tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = string("expand_dims_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = string("expand_dims_0")];
+            tensor<int32, [1]> var_91_axes_0 = const()[name = string("op_91_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_91 = expand_dims(axes = var_91_axes_0, x = mel_length)[name = string("op_91")];
+            tensor<bool, [1, ?]> mask = greater_equal(x = expand_dims_0, y = var_91)[name = string("mask")];
+            tensor<int32, [1]> var_93_axes_0 = const()[name = string("op_93_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, ?]> var_93 = expand_dims(axes = var_93_axes_0, x = mask)[name = string("op_93")];
+            tensor<fp16, [1, 128, ?]> processed_signal_cast_fp16 = select(a = var_16_to_fp16, b = x_cast_fp16, cond = var_93)[name = string("processed_signal_cast_fp16")];
+            string processed_signal_cast_fp16_to_fp32_dtype_0 = const()[name = string("processed_signal_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 128, ?]> mel = cast(dtype = processed_signal_cast_fp16_to_fp32_dtype_0, x = processed_signal_cast_fp16)[name = string("cast_9")];
+        } -> (mel, mel_length);
+}

multilingual/1120ms/preprocessor.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3
+size 592384

multilingual/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b636481110d0a364e9d3379470371c0eda62de5eeac1e58ed3e6371c26c14cb3
+size 15878

multilingual/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3
+size 592384