manifest: declare the audio frontend, transducer decode and driver

parakeet-tdt-0-6b-v3.json

@@ -1,9 +1,16 @@
 {
   "schema": 1,
   "name": "parakeet-tdt-0-6b-v3",
-  "source": "nvidia/parakeet-tdt-0.6b-v3",
   "family": "asr",
+  "source": "nvidia/parakeet-tdt-0.6b-v3",
+  "comment": "",
+  "target": {
+    "backend": "coreml",
+    "min_os": "macOS15",
+    "note": "portable .mlpackage; .mlmodelc is compiled on device and cached content-addressed by model hash + OS build + device model"
+  },
   "arch": {
+    "model_type": "parakeet_tdt",
     "enc_hidden": 1024,
     "enc_layers": 24,
     "enc_heads": 8,
@@ -25,28 +32,129 @@
     "decoder_layers": 2,
     "vocab": 8193,
     "blank_token_id": 8192,
+    "durations": [
+      0,
+      1,
+      2,
+      3,
+      4
+    ],
     "max_symbols_per_step": 10,
     "hidden_act": "relu",
     "joint_vocab_out": 8198
   },
   "artifacts": {
     "graphs": {
-      "encoder": {
-        "package": "parakeet-tdt-0-6b-v3_encoder.mlpackage"
+      "parakeet-tdt-0-6b-v3_encoder": {
+        "role": "encoder",
+        "package": "parakeet-tdt-0-6b-v3_encoder.mlpackage",
+        "compiled": "parakeet-tdt-0-6b-v3_encoder.mlmodelc",
+        "seq": 1,
+        "ctx": 0,
+        "precision": "fp16",
+        "compute_units": "cpuAndNeuralEngine",
+        "stateful": false
       },
-      "step": {
-        "package": "parakeet-tdt-0-6b-v3_step.mlpackage"
+      "parakeet-tdt-0-6b-v3_step": {
+        "role": "step",
+        "package": "parakeet-tdt-0-6b-v3_step.mlpackage",
+        "compiled": "parakeet-tdt-0-6b-v3_step.mlmodelc",
+        "seq": 1,
+        "ctx": 0,
+        "precision": "fp16",
+        "compute_units": "cpuAndNeuralEngine",
+        "stateful": false
       }
     },
     "host": {
-      "embed": "parakeet-tdt-0-6b-v3_embed_f16.bin",
-      "tokenizer": "tokenizer.json"
+      "tokenizer": "tokenizer.json",
+      "tokenizer_kind": "sentencepiece",
+      "embed": "parakeet-tdt-0-6b-v3_embed_f16.bin"
     }
   },
-  "runtime": {
-    "compute_units": "cpuAndNeuralEngine",
-    "precision": "fp16",
-    "ctx": null,
-    "seq": 1
+  "audio": {
+    "kind": "log_mel",
+    "sample_rate": 16000,
+    "buckets": [
+      512
+    ],
+    "flavor": "nemo",
+    "n_mels": 128,
+    "normalize": true
+  },
+  "decode": {
+    "pipeline": "transducer",
+    "driver": "ane_transducer_transcribe",
+    "vocab": 8193,
+    "blank_token_id": 8192,
+    "n_durations": 5,
+    "durations": [
+      0,
+      1,
+      2,
+      3,
+      4
+    ],
+    "max_symbols_per_step": 10,
+    "pred_state": [
+      "h0",
+      "c0",
+      "h1",
+      "c1"
+    ]
+  },
+  "state": {
+    "name": "",
+    "kind": "explicit",
+    "tensors": [
+      "h0",
+      "c0",
+      "h1",
+      "c1"
+    ],
+    "shared_by": [],
+    "note": "the prediction network's LSTM state, threaded through the step graph's own I/O"
+  },
+  "success": {
+    "parity": "wer",
+    "wer_max": 0.05,
+    "min_tok_per_s": 0.0
+  },
+  "plan": {
+    "steps": [
+      {
+        "kind": "host",
+        "op": "ane_transducer_transcribe",
+        "args": {
+          "tokenizer": "tokenizer.json",
+          "tokenizer_kind": "sentencepiece",
+          "kind": "log_mel",
+          "flavor": "nemo",
+          "sample_rate": 16000,
+          "n_mels": 128,
+          "normalize": true,
+          "buckets": [
+            512
+          ],
+          "vocab": 8193,
+          "blank_token_id": 8192,
+          "n_durations": 5,
+          "durations": [
+            0,
+            1,
+            2,
+            3,
+            4
+          ],
+          "max_symbols_per_step": 10,
+          "pred_state": [
+            "h0",
+            "c0",
+            "h1",
+            "c1"
+          ]
+        }
+      }
+    ]
   }
-}
+}