Update app.py

app.py

@@ -22,8 +22,8 @@ pipe_is = pipeline(model=MODEL_IS)
 pipe_fo = pipeline(model=MODEL_FO)
 
 
-whp_is = WhisperProcessor.from_pretrained(MODEL_WHIS)
-whm_is = WhisperForConditionalGeneration.from_pretrained(MODEL_WHIS)
+wdevice = "cuda" if torch.cuda.is_available() else "cpu"
+whm_is = WhisperModel(model_size_or_path=MODEL_WHIS, device=wdevice)
 
 
 
@@ -46,13 +46,12 @@ def recc(audio_file,model,processor):
         xcp = processor.batch_decode(pred_ids)
         return xcp[0]
         
-def whrecc(audio_file,wmodel,wprocessor):
+def whrecc(audio_file,lang,wmodel):
     wav = readwav(audio_file)
-    input_features = wprocessor(wav, sampling_rate=16000, return_tensors="pt").input_features
-    predicted_ids = wmodel.generate(input_features)
-    dec = wprocessor.batch_decode(predicted_ids, skip_special_tokens=True,language_id='is')
-    xcp = dec[0]
-    return xcp
+    xcps, info = wmodel.transcribe(audio = audio_file, language = lang, no_repeat_ngram_size = 5)
+    txts = [xtp.text for xcp in xcps]
+    txt = ' '.join(txts)
+    return txt
 
     
 def recis(audio_file):
@@ -69,7 +68,7 @@ def recfo(audio_file):
 
 
 def recwhis(audio_file):
-    wh_output = whrecc(audio_file,whm_is,whp_is)
+    wh_output = whrecc(audio_file,"is",whm_is)
     return(wh_output)
 
 def pick_asrc(au_src):