Fix model.transcribe

Some of the inference functions now return lists of dicts where they previously returned just dicts, causing them to fail when transcribe is called. This makes a trivial fix of just picking the first list option when that occurs.

whisper/transcribe.py

@@ -134,6 +134,11 @@ def transcribe(
     content_frames = mel.shape[-1] - N_FRAMES
     content_duration = float(content_frames * HOP_LENGTH / SAMPLE_RATE)
 
+    def maybe_dereference_list(obj: Union[dict, list[dict]]) -> dict:
+        if isinstance(obj, list):
+            return obj[0]
+        return obj
+
     if decode_options.get("language", None) is None:
         if not model.is_multilingual:
             decode_options["language"] = "en"
@@ -144,7 +149,7 @@ def transcribe(
                 )
             mel_segment = pad_or_trim(mel, N_FRAMES).to(model.device).to(dtype)
             _, probs = model.detect_language(mel_segment)
-            decode_options["language"] = max(probs, key=probs.get)
+            decode_options["language"] = max(maybe_dereference_list(probs), key=maybe_dereference_list(probs).get)
             if verbose is not None:
                 print(
                     f"Detected language: {LANGUAGES[decode_options['language']].title()}"
@@ -192,7 +197,7 @@ def decode_with_fallback(segment: torch.Tensor) -> DecodingResult:
                 kwargs.pop("best_of", None)
 
             options = DecodingOptions(**kwargs, temperature=t)
-            decode_result = model.decode(segment, options)
+            decode_result = maybe_dereference_list(model.decode(segment, options))
 
             needs_fallback = False
             if (