Update text_to_speech.py

text_to_speech.py

@@ -1,40 +1,37 @@
 import time
 import torch
-from transformers import set_seed
-from transformers import VitsTokenizer, VitsModel
+import scipy
+from transformers import set_seed, pipeline
 
-def synthesize_facebook(s:str, iso3:str) -> str:
-    ''' 
-    For given text, speak it.
-    
-    Parameters
-    ---------- 
-    s: str
-        The written text.
-    is03:str
-        The ISO-3 code of the text's language.
 
-    Returns
-    ---------- 
-    synth:str
-        The synthesized audio.
-    '''
-    
-    # Ensure replicability
-    set_seed(555) 
-    start_time = time.time()
+def goai_tts(texte, device):
+    """
+    Pour un texte donné, donner le speech en Mooré correspondant
     
-    # Load synthesizer
-    tokenizer = VitsTokenizer.from_pretrained(f"facebook/mms-tts-{iso3}")
-    model = VitsModel.from_pretrained(f"facebook/mms-tts-{iso3}")
-    
-    inputs = tokenizer(text=s, return_tensors="pt")
+    Paramètres
+    ---------- 
+    texte: str
+        Le texte écrit.
+    device: str
+        GPU ou CPU
+        
+    Return
+    ------
+        L'audio synthétisé.
+    """
     
-    # Inference
-    with torch.no_grad():
-       outputs = model(**inputs)
+    ### assurer la reproductibilité
+    set_seed(2024) 
     
-    synth = outputs.waveform[0]    
+    start_time = time.time()
+
+    ### charger le modèle TTS
+    model_id = "anyantudre/mms-tts-mos-V1"
+    synthesiser = pipeline("text-to-speech", model_id, device=device)  # add device=0 if you want to use a GPU
+
+    ### inférence
+    speech = synthesiser(texte)
+    wavfile = scipy.io.wavfile.write("finetuned_output.wav", rate=speech["sampling_rate"], data=speech["audio"][0])
 
-    print("Time elapsed: ", int(time.time() - start_time), " seconds")
-    return synth.numpy()
+    print("Temps écoulé: ", int(time.time() - start_time), " seconds")
+    return wavfile