fix(whisper): ignore empty non-speech token encodings

mlx_audio/stt/models/whisper/whisper.py

@@ -183,10 +183,14 @@ def non_speech_tokens(self) -> Tuple[int, ...]:
 
         miscellaneous = set("♩♪♫♬♭♮♯")
 
-        result = {self.encode(" -")[0], self.encode(" '")[0]}
+        result = set()
+        for seed in (" -", " '"):
+            tokens = self.encode(seed)
+            if tokens:
+                result.add(tokens[0])
         for symbol in symbols + list(miscellaneous):
             for tokens in [self.encode(symbol), self.encode(" " + symbol)]:
-                if len(tokens) == 1 or symbol in miscellaneous:
+                if tokens and (len(tokens) == 1 or symbol in miscellaneous):
                     result.add(tokens[0])
 
         return tuple(sorted(result))

tests/test_whisper_decode_options.py

@@ -4,6 +4,7 @@
 from mlx_audio.stt.models.whisper.decoding import DecodingOptions
 from mlx_audio.stt.models.whisper.whisper import (
     _DECODING_OPTION_NAMES,
+    HFTokenizerWrapper,
     _filter_decode_options,
 )
 
@@ -54,5 +55,33 @@ def test_input_is_not_mutated(self):
         self.assertEqual(options, {"language": "en", "frame_threshold": 25})
 
 
+class TestHFTokenizerWrapper(unittest.TestCase):
+    def test_non_speech_tokens_skips_empty_seed_encodings(self):
+        class EmptySeedTokenizer:
+            def encode(self, text, add_special_tokens=False):
+                if text in {" -", " '"}:
+                    return []
+                return [101]
+
+        tokenizer = HFTokenizerWrapper(EmptySeedTokenizer(), multilingual=False)
+
+        self.assertEqual(tokenizer.non_speech_tokens, (101,))
+
+    def test_non_speech_tokens_skips_empty_miscellaneous_encodings(self):
+        miscellaneous = set("♩♪♫♬♭♮♯")
+
+        class EmptyMiscellaneousTokenizer:
+            def encode(self, text, add_special_tokens=False):
+                if text.strip() in miscellaneous:
+                    return []
+                return [101]
+
+        tokenizer = HFTokenizerWrapper(
+            EmptyMiscellaneousTokenizer(), multilingual=False
+        )
+
+        self.assertEqual(tokenizer.non_speech_tokens, (101,))
+
+
 if __name__ == "__main__":
     unittest.main()