Converted to blocks format to allow for tabs (new microphone tab)

app.py

@@ -64,127 +64,164 @@ def format_time(seconds):
 
 #Convert Video/Audio into 16K wav file
 def preprocessAudio(audioFile):
-    os.system(f"ffmpeg -y -i {audioFile.name} -ar 16000 ./audioToConvert.wav")
+    if isinstance(audioFile, str):  # If audioFile is a string (filepath)
+        os.system(f"ffmpeg -y -i {audioFile} -ar 16000 ./audioToConvert.wav")
+    else:  # If audioFile is an object with a name attribute
+        os.system(f"ffmpeg -y -i {audioFile.name} -ar 16000 ./audioToConvert.wav")
 
 #Transcribe!!!
-def Transcribe(file, batch_size):
-    device = "cuda:0" if torch.cuda.is_available() else "cpu"
-    start_time = time.time()
-    model.load_adapter("amh")
-    processor.tokenizer.set_target_lang("amh")
-
-    preprocessAudio(file)
-    block_size = 30
-
-    transcripts = []
-    speech_segments = []
-
-    stream = librosa.stream(
-        "./audioToConvert.wav",
-        block_length=block_size,
-        frame_length=16000,
-        hop_length=16000
-    )
-
-    model.to(device)
-    print(f"Model loaded to {device}: Entering transcription phase")
-
-    #Code for timestamping
-    encoding_start = 0
-    encoding_end = 0
-    sbv_file = open("subtitle.sbv", "w")
-    transcription_file = open("transcription.txt", "w")
-
-    # Create an empty list to hold batches
-    batch = []
-
-    for speech_segment in stream:
-        if len(speech_segment.shape) > 1:
-            speech_segment = speech_segment[:,0] + speech_segment[:,1]
-
-        # Add the current speech segment to the batch
-        batch.append(speech_segment)
-
-        # If the batch is full, process it
-        if len(batch) == batch_size:
-            # Concatenate all segments in the batch along the time axis
-            input_values = processor(batch, sampling_rate=16_000, return_tensors="pt", padding=True)
-            input_values = input_values.to(device)
-            with torch.no_grad():
-                logits = model(**input_values).logits
-            if len(logits.shape) == 1:
-                logits = logits.unsqueeze(0)
-            beam_search_result = beam_search_decoder(logits.to("cpu"))
-
-            # Transcribe each segment in the batch
-            for i in range(batch_size):
-                transcription = " ".join(beam_search_result[i][0].words).strip()
-                transcripts.append(transcription)
-
-                encoding_end = encoding_start + block_size
-                formatted_start = format_time(encoding_start)
-                formatted_end = format_time(encoding_end)
-                sbv_file.write(f"{formatted_start},{formatted_end}\n")
-                sbv_file.write(f"{transcription}\n\n")
-
-                encoding_start = encoding_end
-
-            # Freeing up memory
-            del input_values
-            del logits
-            del transcription
-            torch.cuda.empty_cache()
-            gc.collect()
-
-            # Clear the batch
-            batch = []
-
-    if batch:
-            # Concatenate all segments in the batch along the time axis
-            input_values = processor(batch, sampling_rate=16_000, return_tensors="pt", padding=True)
-            input_values = input_values.to(device)
-            with torch.no_grad():
-                logits = model(**input_values).logits
-            if len(logits.shape) == 1:
-                logits = logits.unsqueeze(0)
-            beam_search_result = beam_search_decoder(logits.to("cpu"))
-
-            # Transcribe each segment in the batch
-            for i in range(len(batch)):
-                transcription = " ".join(beam_search_result[i][0].words).strip()
-                print(transcription)
-                transcripts.append(transcription)
-
-                encoding_end = encoding_start + block_size
-                formatted_start = format_time(encoding_start)
-                formatted_end = format_time(encoding_end)
-                sbv_file.write(f"{formatted_start},{formatted_end}\n")
-                sbv_file.write(f"{transcription}\n\n")
-
-                encoding_start = encoding_end
-
-            # Freeing up memory
-            del input_values
-            del logits
-            del transcription
-            torch.cuda.empty_cache()
-            gc.collect()
-
-    # Join all transcripts into a single transcript
-    transcript = ' '.join(transcripts)
-    transcription_file.write(f"{transcript}")
-    sbv_file.close()
-    transcription_file.close()
-
-    end_time = time.time()
-    print(f"The script ran for {end_time - start_time} seconds.")
-    return(["./subtitle.sbv", "./transcription.txt"])
+def Transcribe(file):
+    try:
+        device = "cuda:0" if torch.cuda.is_available() else "cpu"
+        start_time = time.time()
+        model.load_adapter("amh")
+        processor.tokenizer.set_target_lang("amh")
+
+        preprocessAudio(file)
+        block_size = 30
+        batch_size = 1
+
+        transcripts = []
+        speech_segments = []
+
+        stream = librosa.stream(
+            "./audioToConvert.wav",
+            block_length=block_size,
+            frame_length=16000,
+            hop_length=16000
+        )
+
+        model.to(device)
+        print(f"Model loaded to {device}: Entering transcription phase")
+
+        #Code for timestamping
+        encoding_start = 0
+        encoding_end = 0
+        sbv_file = open("subtitle.sbv", "w")
+        transcription_file = open("transcription.txt", "w")
+
+        # Create an empty list to hold batches
+        batch = []
+
+        for speech_segment in stream:
+            if len(speech_segment.shape) > 1:
+                speech_segment = speech_segment[:,0] + speech_segment[:,1]
+
+            # Add the current speech segment to the batch
+            batch.append(speech_segment)
+
+            # If the batch is full, process it
+            if len(batch) == batch_size:
+                # Concatenate all segments in the batch along the time axis
+                input_values = processor(batch, sampling_rate=16_000, return_tensors="pt", padding=True)
+                input_values = input_values.to(device)
+                with torch.no_grad():
+                    logits = model(**input_values).logits
+                if len(logits.shape) == 1:
+                    logits = logits.unsqueeze(0)
+                beam_search_result = beam_search_decoder(logits.to("cpu"))
+
+                # Transcribe each segment in the batch
+                for i in range(batch_size):
+                    transcription = " ".join(beam_search_result[i][0].words).strip()
+                    transcripts.append(transcription)
+
+                    encoding_end = encoding_start + block_size
+                    formatted_start = format_time(encoding_start)
+                    formatted_end = format_time(encoding_end)
+                    sbv_file.write(f"{formatted_start},{formatted_end}\n")
+                    sbv_file.write(f"{transcription}\n\n")
+
+                    encoding_start = encoding_end
+
+                # Freeing up memory
+                del input_values
+                del logits
+                del transcription
+                torch.cuda.empty_cache()
+                gc.collect()
+
+                # Clear the batch
+                batch = []
+
+        if batch:
+                # Concatenate all segments in the batch along the time axis
+                input_values = processor(batch, sampling_rate=16_000, return_tensors="pt", padding=True)
+                input_values = input_values.to(device)
+                with torch.no_grad():
+                    logits = model(**input_values).logits
+                if len(logits.shape) == 1:
+                    logits = logits.unsqueeze(0)
+                beam_search_result = beam_search_decoder(logits.to("cpu"))
+
+                # Transcribe each segment in the batch
+                for i in range(len(batch)):
+                    transcription = " ".join(beam_search_result[i][0].words).strip()
+                    print(transcription)
+                    transcripts.append(transcription)
+
+                    encoding_end = encoding_start + block_size
+                    formatted_start = format_time(encoding_start)
+                    formatted_end = format_time(encoding_end)
+                    sbv_file.write(f"{formatted_start},{formatted_end}\n")
+                    sbv_file.write(f"{transcription}\n\n")
+
+                    encoding_start = encoding_end
+
+                # Freeing up memory
+                del input_values
+                del logits
+                del transcription
+                torch.cuda.empty_cache()
+                gc.collect()
+
+        # Join all transcripts into a single transcript
+        transcript = ' '.join(transcripts)
+        transcription_file.write(f"{transcript}")
+        sbv_file.close()
+        transcription_file.close()
+
+        end_time = time.time()
+        print(f"The script ran for {end_time - start_time} seconds.")
+        return(["./subtitle.sbv", "./transcription.txt"])
+    except Exception as e:
+        error_log = open("error_log.txt", "w")
+        error_log.write(f"Exception occurred: {e}")
+        error_log.close()
+
+demo = gr.Blocks()
+
+with demo:
+    gr.Markdown(
+    """
+    <div align="center">
     
-demo = gr.Interface(fn=Transcribe, inputs=[gr.File(label="Upload an audio file of Amharic content"), gr.Slider(0, 25, value=4, step=1, label="batch size", info="Approximately .5GB per batch")], 
-                    outputs=gr.File(label="Download .sbv transcription", file_count="multiple"),
-                    title="Amharic Audio Transcription",
-                    description="This application uses Meta MMS and an Amharic kenLM model to transcribe Amharic Audio files of arbitrary length into .sbv and .txt files. Upload an Amharic audio file and get your transcription! \n(Note: Transcription quality is quite low, you should review and edit transcriptions before making them publicly available)"
-                    )
+    # Amharic Audio Transcription
+
+    This application uses Meta MMS and an Amharic kenLM model to transcribe Amharic Audio files of arbitrary length into .sbv and .txt files. Upload an Amharic audio file and get your transcription!
+
+    ### (Note: Transcription quality is quite low, you should review and edit transcriptions before making them publicly available)
+    </div>
+    """)
+    with gr.Tabs():
+        with gr.TabItem("From File"):
+            with gr.Row():
+                file_input = gr.File(label="Upload an audio file of Amharic Content")
+                file_output = gr.Files(label="Download output files")
+            file_button = gr.Button("Submit")
+        with gr.TabItem("From Microphone"):
+            with gr.Row():
+                microphone_input = gr.Audio(type="filepath", source="microphone")
+                microphone_output = gr.Files(label="Download output files")
+            microphone_button = gr.Button("Submit")
+            
+    file_button.click(Transcribe, inputs=file_input, outputs=file_output)
+    microphone_button.click(Transcribe, inputs=microphone_input, outputs=microphone_output)
+        
+# demo = gr.Interface(fn=Transcribe, inputs=[gr.File(label="Upload an audio file of Amharic content"), gr.Slider(0, 25, value=4, step=1, label="batch size", info="Approximately .5GB per batch")], 
+#                     outputs=gr.File(label="Download .sbv transcription", file_count="multiple"),
+#                     title="Amharic Audio Transcription",
+#                     description="This application uses Meta MMS and an Amharic kenLM model to transcribe Amharic Audio files of arbitrary length into .sbv and .txt files. Upload an Amharic audio file and get your transcription! \n(Note: Transcription quality is quite low, you should review and edit transcriptions before making them publicly available)"
+#                     )
 
 demo.launch()
-