Create handler.py

handler.py

@@ -0,0 +1,35 @@
+import torch
+import soundfile as sf
+import numpy as np
+from transformers import CsmForConditionalGeneration, AutoProcessor
+
+class EndpointHandler:
+    def __init__(self, model_path: str = "jsbeaudry/sesame-creole-tts"):
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+        self.device = device
+        self.processor = AutoProcessor.from_pretrained(model_path)
+        self.model = CsmForConditionalGeneration.from_pretrained(model_path, device_map=device)
+
+    def __call__(self, data: dict) -> dict:
+        text = data.get("inputs", "[0]Bonjou tout moun koman nou ye?")
+        sampling_rate = data.get("sampling_rate", 24000)
+
+        # Prepare input
+        inputs = self.processor(text, add_special_tokens=True).to(self.device)
+
+        # Generate audio
+        output = self.model.generate(**inputs, output_audio=True)
+        audio_tensor = output[0].to(torch.float32).cpu().numpy()
+
+        # Return audio as base64-encoded WAV (binary isn't supported directly in response)
+        import io, base64
+        buffer = io.BytesIO()
+        sf.write(buffer, audio_tensor, sampling_rate, format="WAV")
+        buffer.seek(0)
+        audio_base64 = base64.b64encode(buffer.read()).decode("utf-8")
+
+        return {
+            "text": text,
+            "audio_base64": audio_base64,
+            "sampling_rate": sampling_rate,
+        }