Update app.py

app.py

@@ -21,13 +21,25 @@ import torch
 
 # is cuda available?
 
-from easynmt import EasyNMT
-translation_model = EasyNMT('m2m_100_418M', max_new_tokens=60)
 
-asr_model = whisper.load_model("base")
+num_cores = psutil.cpu_count()
+os.environ["OMP_NUM_THREADS"] = f"{num_cores}"
+headers = {'Authorization': os.environ['DeepL_API_KEY']}
+
+device = "cpu"#torch.device("cuda" if torch.cuda.is_available() else "cpu")
+print("DEVICE IS: ")
+print(device)
+
+asr_model_base = whisper.load_model("base", device=device)
+asr_model_small = whisper.load_model("small", device=device)
+whisper_models = {
+    'base': asr_model_base,
+    'small': asr_model_small
+}
+
 transcribe_options = dict(beam_size=3, best_of=3, without_timestamps=False)
 
-translation_models = {
+source_languages  = {
 "Afrikaans":"af",
 "Amharic":"am",
 "Arabic":"ar",
@@ -127,15 +139,44 @@ translation_models = {
 "Yiddish":"yi",
 "Yoruba":"yo",
 "Chinese":"zh",
-"Zulu":"zu"
+"Zulu":"zu",
+"Let the model analyze": "Let the model analyze"
+}
+
+DeepL_language_codes_for_translation = {
+"Bulgarian": "BG",
+"Czech": "CS",
+"Danish": "DA",
+"German": "DE",
+"Greek": "EL",
+"English": "EN",
+"Spanish": "ES",
+"Estonian": "ET",
+"Finnish": "FI",
+"French": "FR",
+"Hungarian": "HU",
+"Indonesian": "ID",
+"Italian": "IT",
+"Japanese": "JA",
+"Lithuanian": "LT",
+"Latvian": "LV",
+"Dutch": "NL",
+"Polish": "PL",
+"Portuguese": "PT",
+"Romanian": "RO",
+"Russian": "RU",
+"Slovak": "SK",
+"Slovenian": "SL",
+"Swedish": "SV",
+"Turkish": "TR",
+"Ukrainian": "UK",
+"Chinese": "ZH"
 }
 
-translation_models_list = [key[0] for key in translation_models.items()]
 
+source_language_list = [key[0] for key in source_languages.items()]
+translation_models_list = [key[0] for key in DeepL_language_codes_for_translation.items()]
 
-device = "cpu"#torch.device("cuda" if torch.cuda.is_available() else "cpu")
-print("DEVICE IS: ")
-print(device)
   
 videos_out_path = Path("./videos_out")
 videos_out_path.mkdir(parents=True, exist_ok=True)
@@ -148,18 +189,19 @@ def get_youtube(video_url):
     
     return abs_video_path
 
-async def speech_to_text(video_file_path, selected_translation_lang):
+async def speech_to_text(video_file_path, selected_translation_lang, whisper_model):
     """
     # Youtube with translated subtitles using OpenAI Whisper and Opus-MT models.
     # Currently supports only English audio
     This space allows you to:
     1. Download youtube video with a given url
     2. Watch it in the first video component
-    3. Run automatic speech recognition on the video using Whisper
-    4. Translate the recognized transcriptions to Finnish, Swedish, Danish
+    3. Run automatic speech recognition on the video using fast Whisper models
+    4. Translate the recognized transcriptions to 26 languages supported by deepL (If source language not supported this will return original transciption)
     5. Burn the translations to the original video and watch the video in the 2nd video component
     
-    Speech Recognition is based on OpenAI Whisper https://github.com/openai/whisper
+    Speech Recognition is based on models from OpenAI Whisper https://github.com/openai/whisper
+    This space is using c++ implementation by https://github.com/ggerganov/whisper.cpp
     """
     
     if(video_file_path == None):
@@ -193,20 +235,48 @@ async def speech_to_text(video_file_path, selected_translation_lang):
                             }
             df = df.append(new_row, ignore_index=True)
         
-        if selected_translation_lang is None:
-                    selected_translation_lang = 'Finnish'
-        
-        sentences = df['text']
-        df['translation'] = translation_model.translate(sentences, target_lang=translation_models.get(selected_translation_lang), max_new_tokens = 50)
-        
-    
-        print('After translation to target language \n')
- 
         return (df)
     except Exception as e:
         raise RuntimeError("Error Running inference with local model", e)
 
 
+
+def translate_transcriptions(df, selected_translation_lang_2):
+    if selected_translation_lang_2 is None:
+            selected_translation_lang_2 = 'English'
+    df.reset_index(inplace=True)
+    
+    print("start_translation")
+    translations = []
+    
+    
+
+    text_combined = ""
+    for i, sentence in enumerate(df['text']):
+        if i == 0:
+            text_combined = sentence
+        else:
+            text_combined = text_combined + '\n' + sentence
+
+    data = {'text': text_combined,
+    'tag_spitting': 'xml',
+    'target_lang': DeepL_language_codes_for_translation.get(selected_translation_lang_2)
+           }
+    try:
+        response = requests.post('https://api-free.deepl.com/v2/translate', headers=headers, data=data)
+    
+        # Print the response from the server
+        translated_sentences = json.loads(response.text)
+        translated_sentences = translated_sentences['translations'][0]['text'].split('\n')
+        df['translation'] = translated_sentences
+    except Exception as e:
+        print(e)
+        df['translation'] = df['text']
+        
+    print("translations done")
+
+    return df
+
 def create_srt_and_burn(df, video_in):
     
     print("Starting creation of video wit srt")
@@ -286,10 +356,12 @@ video_out = gr.Video(label="Video Out", mirror_webcam=False)
 
 
 df_init = pd.DataFrame(columns=['start','end','text','translation'])
-selected_translation_lang = gr.Dropdown(choices=translation_models_list, type="value", value="English", label="Language to translate transcriptions to", interactive=True)
-
-transcription_df = gr.DataFrame(value=df_init,label="Transcription dataframe", row_count=(0, "dynamic"), max_rows = 10)
+selected_source_lang = gr.Dropdown(choices=source_language_list, type="value", value="Let the model analyze", label="Spoken language in video", interactive=True)
+selected_translation_lang_2 = gr.Dropdown(choices=translation_models_list, type="value", value="English", label="In which language you want the transcriptions?", interactive=True)
+selected_whisper_model = gr.Dropdown(choices=whisper_models, type="value", value="base", label="Selected Whisper model", interactive=True)
 
+transcription_df = gr.DataFrame(value=df_init,label="Transcription dataframe", row_count=(0, "dynamic"), max_rows = 10, wrap=True, overflow_row_behaviour='paginate')
+transcription_and_translation_df = gr.DataFrame(value=df_init,label="Transcription and translation dataframe", max_rows = 10, wrap=True, overflow_row_behaviour='paginate')
 
 demo = gr.Blocks(css='''
 #cut_btn, #reset_btn { align-self:stretch; }
@@ -306,14 +378,14 @@ with demo:
             ### This space allows you to: 
             ##### 1. Download youtube video with a given URL
             ##### 2. Watch it in the first video component
-            ##### 3. Run automatic speech recognition on the video using Whisper (Please remember to select translation language)
-            ##### 4. Translate the recognized transcriptions to Finnish, Swedish, Danish
+            ##### 3. Run automatic speech recognition on the video using Whisper
+            ##### 4. Translate the recognized transcriptions to 26 languages supported by deepL
             ##### 5. Burn the translations to the original video and watch the video in the 2nd video component
             ''')
             
         with gr.Column():
             gr.Markdown('''
-            ### 1. Insert Youtube URL below (Some examples below which I suggest to use for first tests)
+            ### 1. Insert Youtube URL below. Some test videos below:
             ##### 1. https://www.youtube.com/watch?v=nlMuHtV82q8&ab_channel=NothingforSale24
             ##### 2. https://www.youtube.com/watch?v=JzPfMbG1vrE&ab_channel=ExplainerVideosByLauren
             ##### 3. https://www.youtube.com/watch?v=S68vvV0kod8&ab_channel=Pearl-CohnTelevision
@@ -334,20 +406,17 @@ with demo:
             with gr.Column():
                 gr.Markdown('''
                 ##### Here you can start the transcription and translation process.
-                ##### Be aware that processing will last for a while (35 second video took around 20 seconds in my testing)
+                ##### Be aware that processing will last some time. With base model it is around 3x speed 
                 ''')
-            transcribe_btn = gr.Button("Step 2. Transcribe and translate audio")
-            
-            transcribe_btn.click(speech_to_text, [video_in, selected_translation_lang], transcription_df)
-            
-    with gr.Row():
-        with gr.Column():
-            selected_translation_lang.render()
+            selected_source_lang.render()
+            selected_whisper_model.render()
+            transcribe_btn = gr.Button("Step 2. Transcribe audio")
+            transcribe_btn.click(speech_to_text, [video_in, selected_source_lang, selected_whisper_model], transcription_df)
+
             
     with gr.Row():
         gr.Markdown('''
-        ##### Here you will get transcription and translation output
-        ##### If you see error please remember to select translation language
+        ##### Here you will get transcription  output
         ##### ''')
 
     with gr.Row():
@@ -356,12 +425,25 @@ with demo:
             
     with gr.Row():
         with gr.Column():
-            translate_and_make_srt_btn = gr.Button("Step 3. Create and burn srt to video")
+            gr.Markdown('''
+            ##### Here you will get translated transcriptions. 
+            ##### Please remember to select target language
+            ##### ''')
+            selected_translation_lang_2.render()
+            translate_transcriptions_button = gr.Button("Step 3. Translate transcription")
+            translate_transcriptions_button.click(translate_transcriptions, [transcription_df, selected_translation_lang_2], transcription_and_translation_df)
+            transcription_and_translation_df.render()
+            
+    with gr.Row():
+        with gr.Column():
+            gr.Markdown('''
+            ##### Now press the Step 4. Button to create output video with translated transcriptions
+            ##### ''')
+            translate_and_make_srt_btn = gr.Button("Step 4. Create and burn srt to video")
             print(video_in)
-            translate_and_make_srt_btn.click(create_srt_and_burn, [transcription_df,video_in], [
+            translate_and_make_srt_btn.click(create_srt_and_burn, [transcription_and_translation_df,video_in], [
                 video_out])
             video_out.render()
 
                 
-if __name__ == "__main__":
-    demo.queue().launch(debug=True, share=False, enable_queue=True)
+demo.launch()