app commit with wav2vec-base-960h v3

app.py

@@ -1,44 +1,66 @@
-import soundfile as sf
+#References: 1. https://www.kdnuggets.com/2021/03/speech-text-wav2vec.html 
+            #2. https://www.youtube.com/watch?v=4CoVcsxZphE 
+            #3. https://www.analyticsvidhya.com/blog/2021/02/hugging-face-introduces-the-first-automatic-speech-recognition-model-wav2vec2/ 
+
+#Importing all the necessary packages
+import nltk
+import librosa
 import torch
-from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
-from transformers import M2M100ForConditionalGeneration, M2M100Tokenizer
 import gradio as gr
-import sox
-
-def convert(inputfile, outfile):
-    sox_tfm = sox.Transformer()
-    sox_tfm.set_output_format(
-        file_type="wav", channels=1, encoding="signed-integer", rate=16000, bits=16
-    )
-    sox_tfm.build(inputfile, outfile)
-    
-    
-    
-processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")
-model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h")
-
-def parse_transcription(wav_file):
-    filename = wav_file.name.split('.')[0]
-    convert(wav_file.name, filename + "16k.wav")
-    speech, _ = sf.read(filename + "16k.wav")
-    input_values = processor(speech, sampling_rate=16_000, return_tensors="pt").input_values
-    logits = model(input_values).logits
-    predicted_ids = torch.argmax(logits, dim=-1)
-    transcription = processor.decode(predicted_ids[0], skip_special_tokens=True)
-    return transcription,
-
-
-
-output1 = gr.outputs.Textbox(label="Transcription in English: ")
-output2 = gr.outputs.Textbox(label="Validated Transcription in English")
-    
-input_ = gr.inputs.Audio(source="microphone", type="file") 
-#gr.Interface(parse_transcription, inputs = input_,  outputs="text", 
-#             analytics_enabled=False, show_tips=False, enable_queue=True).launch(inline=False);
-             
-gr.Interface(parse_transcription, inputs = input_,  outputs=[output1, output2], analytics_enabled=False, 
-                                                                            show_tips=False, 
-                                                                            theme='huggingface',
-                                                                            layout='vertical',
-                                                                            title="Piecurus Test on Speech Transcription",
-                                                                            description="This is a live demo for Speech to Text Translation. Models used: facebook/wav2vec2-base-960h", enable_queue=True).launch( inline=False)
+from transformers import Wav2Vec2Tokenizer, Wav2Vec2ForCTC
+nltk.download("punkt")
+
+#Loading the model and the tokenizer
+model_name = "facebook/wav2vec2-base-960h"
+tokenizer = Wav2Vec2Tokenizer.from_pretrained(model_name)
+model = Wav2Vec2ForCTC.from_pretrained(model_name)
+
+
+def load_data(input_file):
+  
+  """ Function for resampling to ensure that the speech input is sampled at 16KHz.
+  """
+  #read the file
+  speech, sample_rate = librosa.load(input_file)
+  #make it 1-D
+  if len(speech.shape) > 1: 
+      speech = speech[:,0] + speech[:,1]
+  #Resampling at 16KHz since wav2vec2-base-960h is pretrained and fine-tuned on speech audio sampled at 16 KHz.
+  if sample_rate !=16000:
+    speech = librosa.resample(speech, sample_rate,16000)
+  return speech
+  
+  
+
+def correct_casing(input_sentence):
+  """ This function is for correcting the casing of the generated transcribed text
+  """
+  sentences = nltk.sent_tokenize(input_sentence)
+  return (' '.join([s.replace(s[0],s[0].capitalize(),1) for s in sentences]))
+  
+
+
+def asr_transcript(input_file):
+  """This function generates transcripts for the provided audio input
+  """
+  speech = load_data(input_file)
+  
+  #Tokenize
+  input_values = tokenizer(speech, return_tensors="pt").input_values
+  #Take logits
+  logits = model(input_values).logits
+  #Take argmax
+  predicted_ids = torch.argmax(logits, dim=-1)
+  #Get the words from predicted word ids
+  transcription = tokenizer.decode(predicted_ids[0])
+  #Output is all upper case
+  transcription = correct_casing(transcription.lower())
+  return transcription
+  
+
+gr.Interface(asr_transcript,
+             inputs = gr.inputs.Audio(source="microphone", type="filepath", optional=True, label="Please record your voice"),
+             outputs = gr.outputs.Textbox(label="Output Text"),
+             title="ASR using Wav2Vec 2.0",
+             description = "This application displays transcribed text for given audio input",
+             examples = [["Test_File1.wav"], ["Test_File2.wav"], ["Test_File3.wav"]], theme="grass").launch()