Update app.py

app.py

@@ -3,15 +3,62 @@ from transformers import RobertaForQuestionAnswering
 from transformers import BertForQuestionAnswering
 from transformers import AutoTokenizer
 from transformers import pipeline
+import soundfile as sf
+import torch
+from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
+import sox
+import subprocess
+
+
+def read_file_and_process(wav_file):
+    filename = wav_file.split('.')[0]
+    filename_16k = filename + "16k.wav"
+    resampler(wav_file, filename_16k)
+    speech, _ = sf.read(filename_16k)
+    inputs = processor(speech, sampling_rate=16_000, return_tensors="pt", padding=True)
+    
+    return inputs
+
+
+def resampler(input_file_path, output_file_path):
+    command = (
+        f"ffmpeg -hide_banner -loglevel panic -i {input_file_path} -ar 16000 -ac 1 -bits_per_raw_sample 16 -vn "
+        f"{output_file_path}"
+    )
+    subprocess.call(command, shell=True)
+
+
+def parse_transcription(logits):
+    predicted_ids = torch.argmax(logits, dim=-1)
+    transcription = processor.decode(predicted_ids[0], skip_special_tokens=True)
+    return transcription
+
+
+def parse(wav_file):
+    input_values = read_file_and_process(wav_file)
+    with torch.no_grad():
+        logits = model(**input_values).logits
+    user_question = parse_transcription(logits)
+    return user_question
+
+
+model_id = "jonatasgrosman/wav2vec2-large-xlsr-53-persian"
+processor = Wav2Vec2Processor.from_pretrained(model_id)
+model = Wav2Vec2ForCTC.from_pretrained(model_id)
 
 model1 = RobertaForQuestionAnswering.from_pretrained("pedramyazdipoor/persian_xlm_roberta_large")
 tokenizer1 = AutoTokenizer.from_pretrained("pedramyazdipoor/persian_xlm_roberta_large")
 
+
 roberta_large = pipeline(task='question-answering', model=model1, tokenizer=tokenizer1)
 
-def Q_A(question, context):
-  answer_pedram = roberta_large({"question":question, "context":context})['answer']
-  return answer_pedram
+def Q_A(text=None, audio=None, context):
+    if text is None:
+        question = parse(audio)
+    elif audio is None:
+        question = text
+    answer_pedram = roberta_large({"question":question, "context":context})['answer']
+    return answer_pedram
 
 
 
@@ -23,6 +70,10 @@ article = "آموزش داده شده با مدل زبانی روبرتا"
 
 demo = gr.Interface(fn=Q_A, # mapping function from input to output
                     inputs=[gr.Textbox(label='پرسش خود را وارد کنید:', show_label=True, text_align='right', lines=2),
+                            gr.Audio(source="microphone", type="filepath",
+                                     label="لطفا دکمه ضبط صدا را بزنید و شروع به صحبت کنید و بعذ از اتمام صحبت دوباره دکمه ضبط را فشار دهید.",
+                                     show_download_button=True,
+                                     show_edit_button=True,),
                             gr.Textbox(label='متن منبع خود را وارد کنید', show_label=True, text_align='right', lines=8)], # what are the inputs?
                     outputs=gr.Text(show_copy_button=True), # what are the outputs?
                               # our fn has two outputs, therefore we have two outputs