diff --git a/server/voice_generator.py b/server/voice_generator.py
new file mode 100644
index 0000000..81efe2f
--- /dev/null
+++ b/server/voice_generator.py
@@ -0,0 +1,73 @@
+"""XTTS v2 wrapper for voice cloning from non-voice audio samples."""
+
+import os
+import random
+import tempfile
+from pathlib import Path
+
+try:
+    from TTS.api import TTS
+except ImportError:
+    TTS = None  # Tests patch this; real runtime requires the TTS package
+
+from server.config import config
+
+
+class VoiceGenerator:
+    """Generates speech cloned from arbitrary audio samples via XTTS v2."""
+
+    def __init__(
+        self,
+        device: str | None = None,
+        model_name: str | None = None,
+        samples_dir: str | None = None,
+    ):
+        self.device = device or config.device
+        self.model_name = model_name or config.models.xtts_model
+        self.samples_dir = Path(samples_dir or config.samples_dir)
+        if TTS is None:
+            raise RuntimeError(
+                "TTS package is not installed; cannot instantiate VoiceGenerator"
+            )
+        self._tts = TTS(model_name=self.model_name)
+        self._tts.to(self.device)
+
+    def generate(self, text: str, speaker_wav: str | None = None) -> bytes:
+        """Generate speech as WAV bytes. Uses a random clone source if none specified."""
+        if speaker_wav is None:
+            speaker_wav = self.random_clone_source()
+        if speaker_wav is None:
+            raise ValueError("No speaker WAV provided and no samples available")
+
+        # XTTS writes to file, so use a temp file
+        tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
+        tmp.close()
+        try:
+            self._tts.tts_to_file(
+                text=text,
+                speaker_wav=speaker_wav,
+                language=config.models.xtts_language,
+                file_path=tmp.name,
+            )
+            with open(tmp.name, "rb") as f:
+                return f.read()
+        finally:
+            try:
+                os.unlink(tmp.name)
+            except OSError:
+                pass
+
+    def list_clone_sources(self) -> list[str]:
+        """List all WAV files in the samples directory."""
+        if not self.samples_dir.is_dir():
+            return []
+        return [
+            str(p) for p in sorted(self.samples_dir.glob("*.wav"))
+        ]
+
+    def random_clone_source(self) -> str | None:
+        """Pick a random clone source WAV file."""
+        sources = self.list_clone_sources()
+        if not sources:
+            return None
+        return random.choice(sources)
diff --git a/tests/test_voice_generator.py b/tests/test_voice_generator.py
new file mode 100644
index 0000000..969d826
--- /dev/null
+++ b/tests/test_voice_generator.py
@@ -0,0 +1,90 @@
+import os
+import tempfile
+import wave
+from unittest.mock import MagicMock, patch
+
+from server.voice_generator import VoiceGenerator
+
+
+class TestVoiceGenerator:
+    @patch("server.voice_generator.TTS")
+    def test_init_loads_model(self, mock_tts_cls):
+        """Generator loads the XTTS v2 model on init."""
+        mock_tts = MagicMock()
+        mock_tts_cls.return_value = mock_tts
+
+        gen = VoiceGenerator(device="cpu")
+        mock_tts_cls.assert_called_once()
+
+    @patch("server.voice_generator.TTS")
+    def test_generate_returns_wav_bytes(self, mock_tts_cls):
+        """Generate returns WAV bytes."""
+        mock_tts = MagicMock()
+        mock_tts_cls.return_value = mock_tts
+
+        # Create a real WAV file for the mock to "produce"
+        with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as f:
+            tmp_wav = f.name
+            with wave.open(f, "wb") as wf:
+                wf.setnchannels(1)
+                wf.setsampwidth(2)
+                wf.setframerate(22050)
+                wf.writeframes(b"\x00\x00" * 22050)  # 1 second of silence
+
+        try:
+            # Mock tts_to_file to copy our test WAV
+            def fake_tts_to_file(text, speaker_wav, language, file_path):
+                import shutil
+                shutil.copy2(tmp_wav, file_path)
+
+            mock_tts.tts_to_file = fake_tts_to_file
+
+            gen = VoiceGenerator(device="cpu")
+            data = gen.generate("hello", speaker_wav=tmp_wav)
+            assert isinstance(data, bytes)
+            assert len(data) > 0
+        finally:
+            os.unlink(tmp_wav)
+
+    @patch("server.voice_generator.TTS")
+    def test_list_clone_sources(self, mock_tts_cls):
+        """Lists available clone source files."""
+        mock_tts = MagicMock()
+        mock_tts_cls.return_value = mock_tts
+
+        with tempfile.TemporaryDirectory() as samples_dir:
+            # Create some fake sample files
+            for name in ["dog.wav", "machine.wav", "wind.wav"]:
+                with open(os.path.join(samples_dir, name), "wb") as f:
+                    f.write(b"fake")
+
+            gen = VoiceGenerator(device="cpu", samples_dir=samples_dir)
+            sources = gen.list_clone_sources()
+            assert len(sources) == 3
+            assert all(s.endswith(".wav") for s in sources)
+
+    @patch("server.voice_generator.TTS")
+    def test_random_clone_source(self, mock_tts_cls):
+        """Picks a random clone source from samples directory."""
+        mock_tts = MagicMock()
+        mock_tts_cls.return_value = mock_tts
+
+        with tempfile.TemporaryDirectory() as samples_dir:
+            for name in ["a.wav", "b.wav", "c.wav"]:
+                with open(os.path.join(samples_dir, name), "wb") as f:
+                    f.write(b"fake")
+
+            gen = VoiceGenerator(device="cpu", samples_dir=samples_dir)
+            source = gen.random_clone_source()
+            assert source is not None
+            assert source.endswith(".wav")
+
+    @patch("server.voice_generator.TTS")
+    def test_empty_samples_dir(self, mock_tts_cls):
+        """Empty samples dir returns None for random source."""
+        mock_tts = MagicMock()
+        mock_tts_cls.return_value = mock_tts
+
+        with tempfile.TemporaryDirectory() as samples_dir:
+            gen = VoiceGenerator(device="cpu", samples_dir=samples_dir)
+            assert gen.random_clone_source() is None