fix(audio): resample MLX outputs to the declared sample rate

backend/services/tts_backend.py

@@ -1782,7 +1782,14 @@ class MLXAudioBackend(TTSBackend):
                 audio = getattr(result, "audio", result)
                 if hasattr(audio, "numpy"):
                     audio = audio.numpy()
-                pieces.append(np.asarray(audio, dtype=np.float32))
+                audio = np.asarray(audio, dtype=np.float32)
+                sr = getattr(result, "sample_rate", self.sample_rate)
+                if sr != self.sample_rate:
+                    import torchaudio
+                    audio = torchaudio.functional.resample(
+                        torch.from_numpy(audio), sr, self.sample_rate,
+                    ).numpy()
+                pieces.append(audio)
         except TypeError:
             # Some engines don't accept lang_code / ref_audio. Retry with
             # only the universal kwargs.
@@ -1791,7 +1798,14 @@ class MLXAudioBackend(TTSBackend):
                 audio = getattr(result, "audio", result)
                 if hasattr(audio, "numpy"):
                     audio = audio.numpy()
-                pieces.append(np.asarray(audio, dtype=np.float32))
+                audio = np.asarray(audio, dtype=np.float32)
+                sr = getattr(result, "sample_rate", self.sample_rate)
+                if sr != self.sample_rate:
+                    import torchaudio
+                    audio = torchaudio.functional.resample(
+                        torch.from_numpy(audio), sr, self.sample_rate,
+                    ).numpy()
+                pieces.append(audio)
 
         if not pieces:
             raise RuntimeError(f"mlx-audio ({self._model_id}) produced no audio")

docs/engines/mlx-audio.md

@@ -43,7 +43,8 @@ HF repo id. The env var overrides the persisted UI choice.
 
 ## Behaviour notes
 
-- Output is 24 kHz mono for most hosted models.
+- Output is 24 kHz mono. Results from models with a different native rate
+  (such as Dia at 44.1 kHz) are resampled before stitching and export.
 - **Cloning works only with the `csm` model** — it is the only curated model
   confirmed to accept a reference clip. Other models silently ignore
   reference audio, so the engine reports cloning support only when CSM is

tests/test_mlx_audio_instruct_1405.py

@@ -64,6 +64,7 @@ def backend(monkeypatch, tts_backend):
     """An MLXAudioBackend with the model pre-loaded, so no mlx import happens."""
     be = tts_backend.MLXAudioBackend.__new__(tts_backend.MLXAudioBackend)
     be._model_id = "mlx-community/Qwen3-TTS-12Hz-1.7B-VoiceDesign-4bit"
+    be._sr = 24000
     be._model = _FakeModel()
     monkeypatch.setattr(be, "_ensure_loaded", lambda: None)
     return be

tests/test_mlx_audio_sample_rate.py

@@ -0,0 +1,53 @@
+"""MLX results must match the adapter's declared rate before stitching/export."""
+import importlib
+from types import SimpleNamespace
+
+import numpy as np
+import pytest
+
+
+@pytest.fixture
+def backend(monkeypatch):
+    cls = importlib.import_module("services.tts_backend").MLXAudioBackend
+    be = cls.__new__(cls)
+    be._model_id = "mlx-community/Dia-1.6B"
+    be._sr = 24000
+    monkeypatch.setattr(be, "_ensure_loaded", lambda: None)
+    return be
+
+
+def _tone(rate):
+    return np.sin(2 * np.pi * 440 * np.arange(rate) / rate).astype(np.float32)
+
+
+@pytest.mark.parametrize("rate", [24000, 44100, 48000])
+@pytest.mark.parametrize("fallback", [False, True])
+def test_generated_audio_keeps_its_duration_and_pitch(backend, rate, fallback):
+    def generate(**kwargs):
+        if fallback and "voice" in kwargs:
+            raise TypeError("unexpected keyword argument 'voice'")
+        yield SimpleNamespace(audio=_tone(rate), sample_rate=rate)
+
+    backend._model = SimpleNamespace(generate=generate)
+    declared_rate = backend.sample_rate  # callers can read this before generate
+    wav = backend.generate("Hello", voice="speaker")
+
+    assert backend.sample_rate == declared_rate
+    assert wav.shape == (1, declared_rate)
+    spectrum = np.abs(np.fft.rfft(wav[0].numpy()))
+    assert np.fft.rfftfreq(wav.shape[-1], 1 / declared_rate)[spectrum.argmax()] == 440
+
+
+def test_each_piece_is_resampled_before_concatenation(backend):
+    def generate(**kwargs):
+        for rate in (44100, 24000):
+            yield SimpleNamespace(audio=_tone(rate), sample_rate=rate)
+
+    backend._model = SimpleNamespace(generate=generate)
+    assert backend.generate("Two pieces").shape == (1, 48000)
+
+
+def test_raw_audio_without_rate_metadata_keeps_the_existing_contract(backend):
+    audio = _tone(24000)
+    backend._model = SimpleNamespace(generate=lambda **kwargs: iter([audio]))
+    np.testing.assert_array_equal(backend.generate("Hello")[0].numpy(), audio)