try live=True

app.py

@@ -1,28 +1,21 @@
 import gradio as gr
 import os
 import time
-import sys
-import subprocess
 import tempfile
 import requests
 from urllib.parse import urlparse
 from pydub import AudioSegment
+import logging
+import torch
+from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
+import yt_dlp
 
-# Clone and install faster-whisper from GitHub
-# (we should be able to do this in build.sh in a hf space)
-try:
-    subprocess.run(["git", "clone", "https://github.com/SYSTRAN/faster-whisper.git"], check=True)
-    subprocess.run(["pip", "install", "-e", "./faster-whisper"], check=True)
-except subprocess.CalledProcessError as e:
-    print(f"Error during faster-whisper installation: {e}")
-    sys.exit(1)
+logging.basicConfig(level=logging.INFO)
 
-# Add the faster-whisper directory to the Python path
 sys.path.append("./faster-whisper")
+from faster_whisper import WhisperModel, BatchedInferencePipeline
 
-from faster_whisper import WhisperModel
-from faster_whisper.transcribe import BatchedInferencePipeline
-import yt_dlp
+device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
 def download_audio(url, method_choice):
     parsed_url = urlparse(url)
@@ -31,7 +24,6 @@ def download_audio(url, method_choice):
     else:
         return download_direct_audio(url, method_choice)
 
-# Additional YouTube download methods
 def download_youtube_audio(url, method_choice):
     methods = {
         'yt-dlp': youtube_dl_method,
@@ -41,13 +33,12 @@ def download_youtube_audio(url, method_choice):
         'ffmpeg': ffmpeg_method,
         'aria2': aria2_method
     }
-    
     method = methods.get(method_choice, youtube_dl_method)
-    
     try:
         return method(url)
     except Exception as e:
-        return f"Error downloading using {method_choice}: {str(e)}"
+        logging.error(f"Error downloading using {method_choice}: {str(e)}")
+        return None
 
 def youtube_dl_method(url):
     ydl_opts = {
@@ -74,7 +65,6 @@ def pytube_method(url):
     return new_file
 
 def youtube_dl_classic_method(url):
-    # Classic youtube-dl method
     ydl_opts = {
         'format': 'bestaudio/best',
         'postprocessors': [{
@@ -131,8 +121,9 @@ def download_direct_audio(url, method_choice):
             else:
                 raise Exception(f"Failed to download audio from {url}")
         except Exception as e:
-            return f"Error downloading direct audio: {str(e)}"
-        
+            logging.error(f"Error downloading direct audio: {str(e)}")
+            return None
+
 def wget_method(url):
     output_file = tempfile.mktemp(suffix='.mp3')  
     command = ['wget', '-O', output_file, url]
@@ -140,44 +131,43 @@ def wget_method(url):
     return output_file
 
 def trim_audio(audio_path, start_time, end_time):
-    audio = AudioSegment.from_mp3(audio_path)
+    audio = AudioSegment.from_file(audio_path)
     trimmed_audio = audio[start_time*1000:end_time*1000] if end_time else audio[start_time*1000:]
-    trimmed_audio_path = tempfile.mktemp(suffix='.mp3')
-    trimmed_audio.export(trimmed_audio_path, format="mp3")
+    trimmed_audio_path = tempfile.mktemp(suffix='.wav')
+    trimmed_audio.export(trimmed_audio_path, format="wav")
     return trimmed_audio_path
 
+def save_transcription(transcription):
+    file_path = tempfile.mktemp(suffix='.txt')
+    with open(file_path, 'w') as f:
+        f.write(transcription)
+    return file_path
+
 def transcribe_audio(input_source, batch_size, download_method, start_time=None, end_time=None, verbose=False):
     try:
-        # Initialize the model
         model = WhisperModel("cstr/whisper-large-v3-turbo-int8_float32", device="auto", compute_type="int8")
         batched_model = BatchedInferencePipeline(model=model)
 
-        # Handle input source
         if isinstance(input_source, str) and (input_source.startswith('http://') or input_source.startswith('https://')):
-            # It's a URL, download the audio
             audio_path = download_audio(input_source, download_method)
             if audio_path.startswith("Error"):
                 yield f"Error: {audio_path}", "", None
                 return
         else:
-            # It's a local file path
             audio_path = input_source
 
-        # Trim the audio if start_time or end_time is specified
         if start_time is not None or end_time is not None:
             trimmed_audio_path = trim_audio(audio_path, start_time or 0, end_time)
             audio_path = trimmed_audio_path
 
-        # Benchmark transcription time
         start_time_perf = time.time()
         segments, info = batched_model.transcribe(audio_path, batch_size=batch_size, initial_prompt=None)
         end_time_perf = time.time()
 
-        # Show initial metrics as soon as possible
         transcription_time = end_time_perf - start_time_perf
         real_time_factor = info.duration / transcription_time
-        audio_file_size = os.path.getsize(audio_path) / (1024 * 1024)  # Size in MB
-        
+        audio_file_size = os.path.getsize(audio_path) / (1024 * 1024)
+
         metrics_output = (
             f"Language: {info.language}, Probability: {info.language_probability:.2f}\n"
             f"Duration: {info.duration:.2f}s, Duration after VAD: {info.duration_after_vad:.2f}s\n"  
@@ -191,15 +181,13 @@ def transcribe_audio(input_source, batch_size, download_method, start_time=None,
 
         transcription = ""
 
-        # Stream transcription output gradually
         for segment in segments:
             transcription_segment = f"[{segment.start:.2f}s -> {segment.end:.2f}s] {segment.text}\n"
             transcription += transcription_segment
 
-            if verbose:  
+            if verbose:
                 yield metrics_output, transcription, None
 
-        # Final output with download option
         transcription_file = save_transcription(transcription)
         yield metrics_output, transcription, transcription_file
 
@@ -207,7 +195,6 @@ def transcribe_audio(input_source, batch_size, download_method, start_time=None,
         yield f"An error occurred: {str(e)}", "", None
 
     finally:
-        # Clean up downloaded and trimmed files
         if isinstance(input_source, str) and (input_source.startswith('http://') or input_source.startswith('https://')):
             try:
                 os.remove(audio_path)
@@ -219,17 +206,10 @@ def transcribe_audio(input_source, batch_size, download_method, start_time=None,
             except:
                 pass
 
-def save_transcription(transcription):
-    file_path = tempfile.mktemp(suffix='.txt')
-    with open(file_path, 'w') as f:
-        f.write(transcription)
-    return file_path
-
-# Gradio interface
 iface = gr.Interface(
     fn=transcribe_audio,
     inputs=[
-        gr.Textbox(label="Audio Source (Upload, MP3 URL, or YouTube URL)"),
+        gr.Textbox(label="Audio Source (Upload, URL, or YouTube URL)"),
         gr.Slider(minimum=1, maximum=32, step=1, value=16, label="Batch Size"),
         gr.Dropdown(choices=["yt-dlp", "pytube", "youtube-dl", "yt-dlp-alt", "ffmpeg", "aria2", "wget"], label="Download Method", value="yt-dlp"),
         gr.Number(label="Start Time (seconds)", value=0),  
@@ -237,18 +217,19 @@ iface = gr.Interface(
         gr.Checkbox(label="Verbose Output", value=False)
     ],
     outputs=[
-        gr.Textbox(label="Transcription Metrics and Verbose Messages"),
-        gr.Textbox(label="Transcription"),
-        gr.File(label="Download Transcription")  
+        gr.Textbox(label="Transcription Metrics and Verbose Messages", lines=10),
+        gr.Textbox(label="Transcription", lines=10),
+        gr.File(label="Download Transcription")
     ],
-    title="Faster Whisper Multi-Input Transcription",
-    description="Enter an audio file path, MP3 URL, or YouTube URL to transcribe using Faster Whisper (GitHub version). Adjust the batch size and choose a download method.",
+    title="Multi-Model Transcription",
+    description="Transcribe audio using with Whisper.",
     examples=[
         ["https://www.youtube.com/watch?v=daQ_hqA6HDo", 16, "yt-dlp", 0, None, False], 
         ["https://mcdn.podbean.com/mf/web/dir5wty678b6g4vg/HoP_453_-_The_Price_is_Right_-_Law_and_Economics_in_the_Second_Scholastic5yxzh.mp3", 16, "ffmpeg", 0, 300, True],
         ["path/to/local/audio.mp3", 16, "yt-dlp", 60, 180, False]
     ],
-    cache_examples=False  # Prevents automatic processing of examples
+    cache_examples=False,
+    live=True
 )
 
 iface.launch()