fixed bug on max output text len + adding GPU inference

app.py

@@ -12,6 +12,7 @@ import torch
 import gradio as gr
 from transformers import Wav2Vec2Tokenizer, Wav2Vec2ForCTC
 nltk.download("punkt")
+torch_device = 'cuda'
 
 
 # In[ ]:
@@ -22,7 +23,7 @@ model_name = "facebook/wav2vec2-base-960h"
 
 #model_name = "facebook/wav2vec2-large-xlsr-53"
 tokenizer = Wav2Vec2Tokenizer.from_pretrained(model_name)#omdel_name
-model = Wav2Vec2ForCTC.from_pretrained(model_name)
+model = Wav2Vec2ForCTC.from_pretrained(model_name).to(torch_device)
 
 
 # In[ ]:
@@ -81,10 +82,10 @@ def asr_transcript_long(input_file,tokenizer=tokenizer, model=model ):
     # Ensure that the sample rate is 16k
     sample_rate = librosa.get_samplerate(input_file)
 
-    # Stream over 30 seconds chunks rather than load the full file
+    # Stream over 10 seconds chunks rather than load the full file
     stream = librosa.stream(
         input_file,
-        block_length=30,
+        block_length=20, #number of seconds to split the batch
         frame_length=sample_rate, #16000,
         hop_length=sample_rate, #16000
     )
@@ -95,15 +96,15 @@ def asr_transcript_long(input_file,tokenizer=tokenizer, model=model ):
         if sample_rate !=16000:
             speech = librosa.resample(speech, sample_rate,16000)
         input_values = tokenizer(speech, return_tensors="pt").input_values
-        logits = model(input_values).logits
+        logits = model(input_values.to(torch_device)).logits
 
         predicted_ids = torch.argmax(logits, dim=-1)
         transcription = tokenizer.decode(predicted_ids[0])
-        #transcript += correct_sentence(transcription.lower())
+        #transcript += transcription.lower()
         transcript += correct_casing(transcription.lower())
-        transcript += " "
+        #transcript += " "
 
-    return transcript
+    return transcript[:4300]
 
 
 # In[ ]:
@@ -112,8 +113,8 @@ def asr_transcript_long(input_file,tokenizer=tokenizer, model=model ):
 gr.Interface(asr_transcript_long,
              #inputs = gr.inputs.Audio(source="microphone", type="filepath", optional=True, label="Please record your voice"),
              inputs = gr.inputs.Audio(source="upload", type="filepath", optional=True, label="Upload your file here"),
-             outputs = gr.outputs.Textbox(label="Output Text"),
-             title="ASR using Wav2Vec 2.0",
+             outputs = gr.outputs.Textbox(type="str",label="Output Text"),
+             title="Transcript and Translate",
              description = "This application displays transcribed text for given audio input",
              examples = [["Test_File1.wav"], ["Test_File2.wav"], ["Test_File3.wav"]], theme="grass").launch()