Update app.py

app.py

@@ -1,20 +1,17 @@
+import torch  # Add this line
 import gradio as gr
-from transformers import pipeline, AutoTokenizer
-from huggingsound import SpeechRecognitionModel
+from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor, pipeline, AutoTokenizer
 import numpy as np
 import soundfile as sf
 import tempfile
 
-## Load the model for speech recognition
-model = SpeechRecognitionModel("jonatasgrosman/wav2vec2-large-xlsr-53-english")
-
+# Load the models and processors
+asr_model = Wav2Vec2ForCTC.from_pretrained("jonatasgrosman/wav2vec2-large-xlsr-53-english")
+asr_processor = Wav2Vec2Processor.from_pretrained("jonatasgrosman/wav2vec2-large-xlsr-53-english")
 translator = pipeline("text2text-generation", model="dammyogt/damilola-finetuned-NLP-opus-mt-en-ha")
 tts = pipeline("text-to-speech", model="Baghdad99/hausa_voice_tts")
 
-# Define the function to translate speech
 def translate_speech(audio_data_tuple):
-    print(f"Type of audio: {type(audio_data_tuple)}, Value of audio: {audio_data_tuple}")  # Debug line
-
     # Extract the audio data from the tuple
     sample_rate, audio_data = audio_data_tuple
 
@@ -22,29 +19,34 @@ def translate_speech(audio_data_tuple):
     with tempfile.NamedTemporaryFile(suffix=".wav", delete=True) as temp_audio_file:
         sf.write(temp_audio_file.name, audio_data, sample_rate)
 
-        # Use the speech recognition model to transcribe the audio
-        output = model.transcribe([temp_audio_file.name])
-        print(f"Output: {output}")  # Print the output to see what it contains
-    
-    # Extract the transcriptions from the output
-    transcriptions = [item['transcription'] for item in output]
+        # Prepare the input dictionary
+        input_dict = asr_processor(temp_audio_file.name, return_tensors="pt", padding=True)
+
+        # Use the ASR model to get the logits
+        logits = asr_model(input_dict.input_values.to("cpu")).logits
+
+        # Get the predicted IDs
+        pred_ids = torch.argmax(logits, dim=-1)[0]
+
+        # Decode the predicted IDs to get the transcription
+        transcription = asr_processor.decode(pred_ids)
+        print(f"Transcription: {transcription}")  # Print the transcription
 
-    for transcription in transcriptions:
         # Use the translation pipeline to translate the transcription
         translated_text = translator(transcription, return_tensors="pt")
-        print(f"Translated text: {translated_text}")  # Print the translated text to see what it contains
+        print(f"Translated text: {translated_text}")  # Print the translated text
 
         # Check if the translated text contains 'generated_token_ids'
         if 'generated_token_ids' in translated_text[0]:
             # Decode the tokens into text
             translated_text_str = translator.tokenizer.decode(translated_text[0]['generated_token_ids'])
+            print(f"Translated text string: {translated_text_str}")  # Print the translated text string
         else:
             print("The translated text does not contain 'generated_token_ids'")
             return
 
         # Use the text-to-speech pipeline to synthesize the translated text
         synthesised_speech = tts(translated_text_str)
-        print(f"Synthesised speech: {synthesised_speech}")  # Print the synthesised speech to see what it contains
 
         # Check if the synthesised speech contains 'audio'
         if 'audio' in synthesised_speech: