Update audio_model.py

audio_model.py

@@ -1,55 +1,26 @@
-from transformers import AutoProcessor, BlipForConditionalGeneration, AutoTokenizer,SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
+from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
 import librosa
 import numpy as np
 import torch
-
-#CONSTANTS
-speaker_embeddings = {
-    "BDL": "spkemb/cmu_us_bdl_arctic-wav-arctic_a0009.npy",
-    "CLB": "spkemb/cmu_us_clb_arctic-wav-arctic_a0144.npy",
-    "RMS": "spkemb/cmu_us_rms_arctic-wav-arctic_b0353.npy",
-    "SLT": "spkemb/cmu_us_slt_arctic-wav-arctic_a0508.npy",
-}
+from datasets import load_dataset
 
 # Carga el modelo de clasificación de tetxo a audio speech
 checkpoint = "microsoft/speecht5_tts"
 processor = SpeechT5Processor.from_pretrained(checkpoint)
 model = SpeechT5ForTextToSpeech.from_pretrained(checkpoint)
 vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
+embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
+speaker_embeddings = torch.tensor(embeddings_dataset[7440]["xvector"]).unsqueeze(0)
 
 device = "cuda" if torch.cuda.is_available() else "cpu"
 
 ### TEXT TO AUDIO SPEECH  MODEL 2    
 # Define la función que convierte texto en voz
-def text_to_speech(text,speaker):
-    # Genera el audio utilizando el modelo 
-    if len(text.strip()) == 0:
-       return (16000, np.zeros(0).astype(np.int16))
+def synthesize_speech(text):
+    text = cleanup_text(text)
     inputs = processor(text=text, return_tensors="pt")
 
-    # limit input length
-    input_ids = inputs["input_ids"]
-    input_ids = input_ids[..., :model.config.max_text_positions]
-
-    if speaker == "Surprise Me!":
-        # load one of the provided speaker embeddings at random
-        idx = np.random.randint(len(speaker_embeddings))
-        key = list(speaker_embeddings.keys())[idx]
-        speaker_embedding = np.load(speaker_embeddings[key])
-
-        # randomly shuffle the elements
-        np.random.shuffle(speaker_embedding)
-
-        # randomly flip half the values
-        x = (np.random.rand(512) >= 0.5) * 1.0
-        x[x == 0] = -1.0
-        speaker_embedding *= x
-
-        #speaker_embedding = np.random.rand(512).astype(np.float32) * 0.3 - 0.15
-    speaker_embedding = torch.tensor(speaker_embedding).unsqueeze(0)
-
-    speech = model.generate_speech(input_ids, speaker_embedding, vocoder=vocoder)
-
-    speech = (speech.numpy() * 32767).astype(np.int16)
-    return (16000, speech)
+    speech = model.generate_speech(inputs["input_ids"], speaker_embeddings, vocoder=vocoder)
+    
+    return gr.Audio.update(value=(16000, speech.cpu().numpy()))
 ### END TEXT TO AUDIO SPEECH  MODEL 2