Update app.py

app.py

@@ -1,55 +1,83 @@
+import os
+os.system("pip install git+https://github.com/openai/whisper.git")
+
+
 import gradio as gr
-from pytube import YouTube
 import whisper
+import io
+import os
+import numpy as np
+from datetime import datetime
 
-# define function for transcription
-def whisper_transcript(model_size, url, audio_file):
-    if url:
-        link = YouTube(url)
-        source = link.streams.filter(only_audio=True)[0].download(filename="audio.mp4")
+import assets
 
+def sendToWhisper(audio_record, audio_upload, task, models_selected, language_toggle, language_selected, without_timestamps):
+    results = []
+
+    audio = None
+    if audio_record is not None:
+        audio = audio_record
+    elif audio_upload is not None:
+        audio = audio_upload
     else:
-        source = audio_file
+        return [["Invalid input"]*5]
 
-    if model_size.endswith(".en"):
-        language = "english"
+    audio = whisper.load_audio(audio)
+    audio = whisper.pad_or_trim(audio)
+
+    for model_name in models_selected:
+        start = datetime.now()
+        model = whisper.load_model(model_name)
+        mel = whisper.log_mel_spectrogram(audio).to(model.device)
+        options = whisper.DecodingOptions(fp16 = False, without_timestamps=without_timestamps, task=task)
+        if language_toggle:
+            options = whisper.DecodingOptions(fp16 = False, without_timestamps=without_timestamps, task=task, language=language_selected)
+        language = ""
+        prob = 0
+        if model_name in assets.lang_detect:
+            _, probs = model.detect_language(mel)
+            language = max(probs, key=probs.get)
+            prob = probs[language]
+        else:
+            language="en"
+            options = whisper.DecodingOptions(fp16 = False, without_timestamps=without_timestamps, task=task, language="en")
+        output_text = whisper.decode(model, mel, options)
+        results.append([model_name, output_text.text, language, str(prob), str((datetime.now() - start).total_seconds())])
+    return results
+
+avail_models = whisper.available_models()
+
+
+
+with gr.Blocks(css=assets.css) as demo:
+    gr.Markdown("This is a demo to use Open AI's Speech to Text (ASR) Model: Whisper. Learn more about the models here on [Github](https://github.com/openai/whisper/search?q=DecodingOptions&type=) FYI: The larger models take a lot longer to transcribe the text :)")
+    gr.Markdown("Here are sample audio files to try out: [Sample Audio](https://drive.google.com/drive/folders/1qYek06ZVeKr9f5Jf35eqi-9CnjNIp98u?usp=sharing)")
+    gr.Markdown("Built by:[@davidtsong](https://twitter.com/davidtsong)")
+    
+    # with gr.Row():
+    with gr.Column():
+
+        # with gr.Column():
+        gr.Markdown("## Input")
+
+        with gr.Row():
+            audio_record = gr.Audio(source="microphone", label="Audio to transcribe", type="filepath",elem_id="audio_inputs")
+            audio_upload = gr.Audio(source="upload", type="filepath", interactive=True,elem_id="audio_inputs")
+
+        models_selected = gr.CheckboxGroup(avail_models, label="Models to use")
+        with gr.Accordion("Settings", open=False):
+            task = gr.Dropdown(["transcribe", "translate"], label="Task", value="transcribe")
+            language_toggle = gr.Dropdown(["Automatic", "Manual"], label="Language Selection", value="Automatic")
+            language_selected = gr.Dropdown(list(assets.LANGUAGES.keys()), label="Language")
+            without_timestamps = gr.Checkbox(label="Without timestamps",value=True)
+        submit = gr.Button(label="Run")
+        
+    # with gr.Row():
+        # with gr.Column():
+        gr.Markdown("## Output")
+        output = gr.Dataframe(headers=["Model", "Text", "Language", "Language Confidence","Time(s)"], label="Results", wrap=True)
+
+    submit.click(fn=sendToWhisper, inputs=[audio_record, audio_upload, task, models_selected, language_toggle, language_selected, without_timestamps], outputs=output)
+
+demo.launch()   
 
-    else:
-        language = None
-
-    options = whisper.DecodingOptions(without_timestamps=True)
-
-    loaded_model = whisper.load_model(model_size)
-    transcript = loaded_model.transcribe(source, language=language)
-
-    return transcript["text"]
-
-# define Gradio app interface
-gradio_ui = gr.Interface(
-    fn=whisper_transcript,
-    title="Transcribe multi-lingual audio clips with Whisper",
-    description="**How to use**: Select a model, paste in a Youtube link or upload an audio clip, then click submit. If your clip is **100% in English, select models ending in ‘.en’**. If the clip is in other languages, or a mix of languages, select models without ‘.en’",
-    article="**Note**: The larger the model size selected or the longer the audio clip, the more time it would take to process the transcript.",
-    inputs=[
-        gr.Dropdown(
-            label="Select Model",
-            choices=[
-                "tiny.en",
-                "base.en",
-                "small.en",
-                "medium.en",
-                "tiny",
-                "base",
-                "small",
-                "medium",
-                "large",
-            ],
-            value="base",
-        ),
-        gr.Textbox(label="Paste YouTube link here"),
-        gr.Audio(label="Upload Audio File", source="upload", type="filepath"),
-    ],
-    outputs=gr.outputs.Textbox(label="Whisper Transcript"),
-)
-
-gradio_ui.queue().launch()