Return flat mono audio arrays

kittentts/onnx_model.py

@@ -15,6 +15,12 @@ def basic_english_tokenize(text):
     tokens = re.findall(r"\w+|[^\w\s]", text)
     return tokens
 
+
+def mono_audio_array(audio):
+    """Return generated mono audio as a flat samples array."""
+    return np.asarray(audio).squeeze()
+
+
 class TextCleaner:
     def __init__(self, dummy=None):
         _pad = "$"
@@ -154,7 +160,7 @@ def generate_single_chunk(self, text: str, voice: str = "expr-voice-5-m", speed:
         # Trim audio
         audio = outputs[0][..., :-5000]
 
-        return audio
+        return mono_audio_array(audio)
 
     def generate_to_file(self, text: str, output_path: str, voice: str = "expr-voice-5-m", 
                           speed: float = 1.0, sample_rate: int = 24000, clean_text: bool=True) -> None:

tests/test_text_normalization.py

@@ -1,4 +1,7 @@
 import unittest
+import sys
+import types
+from unittest.mock import patch
 
 from kittentts import NormalizedTextResult, normalize_text
 from kittentts.preprocess import chunk_text
@@ -62,5 +65,48 @@ def test_unsupported_locale_fails_explicitly(self):
             normalize_text("Bonjour 2026", locale="fr-FR")
 
 
+class AudioArrayTests(unittest.TestCase):
+    def test_mono_audio_array_flattens_single_channel_output(self):
+        espeakng_loader = types.SimpleNamespace(
+            get_library_path=lambda: "/tmp/libespeak-ng.dylib",
+            get_data_path=lambda: "/tmp/espeak-ng-data",
+        )
+        espeak_wrapper = types.SimpleNamespace(set_library=lambda path: None)
+        phonemizer = types.SimpleNamespace(
+            backend=types.SimpleNamespace(
+                EspeakBackend=object,
+                espeak=types.SimpleNamespace(wrapper=types.SimpleNamespace(EspeakWrapper=espeak_wrapper)),
+            )
+        )
+
+        with patch.dict(
+            sys.modules,
+            {
+                "espeakng_loader": espeakng_loader,
+                "phonemizer": phonemizer,
+                "phonemizer.backend": phonemizer.backend,
+                "phonemizer.backend.espeak": phonemizer.backend.espeak,
+                "phonemizer.backend.espeak.wrapper": phonemizer.backend.espeak.wrapper,
+                "numpy": types.SimpleNamespace(asarray=lambda audio: audio, ndarray=object),
+                "onnxruntime": types.SimpleNamespace(),
+                "soundfile": types.SimpleNamespace(),
+            },
+        ):
+            from kittentts.onnx_model import mono_audio_array
+
+        audio = mono_audio_array(_FakeAudioArray((1, 24000), (24000,)))
+
+        self.assertEqual(audio.shape, (24000,))
+
+
+class _FakeAudioArray:
+    def __init__(self, shape, squeezed_shape):
+        self.shape = shape
+        self._squeezed_shape = squeezed_shape
+
+    def squeeze(self):
+        return _FakeAudioArray(self._squeezed_shape, self._squeezed_shape)
+
+
 if __name__ == "__main__":
     unittest.main()