[codex] Fuse NeMo transcribe diarize actor

models/common/speech.py

@@ -11,6 +11,16 @@
 
 from daft import DataType
 
+# One ASR word timestamp. `text` is the normalized word/token text emitted by
+# the backend, with timestamps in the original source timeline.
+WordStruct = DataType.struct(
+    {
+        "text": DataType.string(),
+        "start": DataType.float64(),
+        "end": DataType.float64(),
+    }
+)
+
 # One transcript line. `speaker` is filled in after diarization (empty until then).
 SegmentStruct = DataType.struct(
     {
@@ -19,6 +29,7 @@
         "end": DataType.float64(),
         "text": DataType.string(),
         "speaker": DataType.string(),
+        "words": DataType.list(WordStruct),
     }
 )
 

models/common/vad.py

@@ -90,7 +90,25 @@ def __init__(
 
         self.torch = torch
         self.device = device or ("cuda" if torch.cuda.is_available() else "cpu")
-        self.model = nemo_asr.models.EncDecFrameClassificationModel.from_pretrained(model_id).eval().to(self.device)
+        self.model = (
+            nemo_asr.models.EncDecFrameClassificationModel.from_pretrained(model_id, strict=False).eval().to(self.device)
+        )
+        self.configure(
+            threshold=threshold,
+            min_speech_duration_ms=min_speech_duration_ms,
+            min_silence_duration_ms=min_silence_duration_ms,
+            speech_pad_ms=speech_pad_ms,
+        )
+
+    def configure(
+        self,
+        *,
+        threshold: float,
+        min_speech_duration_ms: int,
+        min_silence_duration_ms: int,
+        speech_pad_ms: int,
+    ) -> None:
+        """Retune threshold/windowing without reloading the NeMo weights."""
         self.threshold = threshold
         self.min_speech_frames = max(1, min_speech_duration_ms // self.FRAME_MS)
         self.min_silence_frames = max(1, min_silence_duration_ms // self.FRAME_MS)