Update handler.py

handler.py

@@ -1,5 +1,4 @@
 from typing import  Dict, Any,Union
-import librosa
 import tempfile
 import numpy as np
 import torch
@@ -8,6 +7,7 @@ import noisereduce as nr
 from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
 from num2tib.core import convert
 from num2tib.core import convert2text
+import soundfile as sf
 import base64
 import re
 import requests
@@ -50,10 +50,6 @@ replacements = [
     ('╗','')
 ]
 
-
-
-
-
 class EndpointHandler():
     def __init__(self, path=""):
         # load the model
@@ -88,13 +84,24 @@ class EndpointHandler():
         speaker_embedding = torch.tensor(speaker_embedding)
         speech = self.model.generate_speech(input_ids.to('cuda'), speaker_embedding.to('cuda'), vocoder=self.vocoder.to('cuda'))
         speech = nr.reduce_noise(y=speech.to('cpu'), sr=16000)
+        # Create a unique temporary WAV file
         with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as temp_wav_file:
             temp_wav_path = temp_wav_file.name
-            librosa.output.write_wav(temp_wav_path, speech.numpy(), sr=16000)
+            sf.write(temp_wav_path, speech.numpy(), 16000, 'PCM_24')  # Use sf.write to write the WAV file
+
+        # Read the WAV file and encode it as base64
         with open(temp_wav_path, "rb") as wav_file:
             audio_base64 = base64.b64encode(wav_file.read()).decode("utf-8")
+
+        # Clean up the temporary WAV file
         os.remove(temp_wav_path)
+
         return {
             "sample_rate": 16000,
-            "audio_base64": audio_base64,  
-        }
+            "audio_base64": audio_base64,  # Base64-encoded audio data
+        }
+
+
+
+
+