Update app.py

app.py

@@ -1,3 +1,6 @@
+
+
+
 import torch
 
 import gradio as gr
@@ -14,6 +17,9 @@ import os
 import time
 import demucs.api
 
+import tqdm
+
+os.environ["GRADIO_TEMP_DIR"] = "/home/yoach/spaces/tmp"
 
 
 MODEL_NAME = "openai/whisper-large-v3"
@@ -39,30 +45,32 @@ def separate_vocal(path):
     return path
 
     
-
-# def separate_vocal(path, track_name, output_folder, demucs_model_name = "htdemucs_ft"):
-#     
-#   os.system(f"python3 -m demucs.separate --two-stems=vocals -n {demucs_model_name} {path} -o {output_folder}")
-#   
-#   return os.path.join(output_folder, demucs_model_name, track_name, "vocals.wav")
-
-
-def transcribe(inputs_path, task, use_demucs, dataset_name, oauth_token: gr.OAuthToken | None):
+def transcribe(inputs_path, task, use_demucs, dataset_name, oauth_token: gr.OAuthToken | None, progress=gr.Progress(track_tqdm=True)):
     if inputs_path is None:
         raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting your request.")
     
+    pbar = tqdm.tqdm(total=4, desc="Overall progression")
+
     sampling_rate, inputs = wavfile.read(inputs_path) 
 
+    pbar.update(1)
+    pbar.set_description("Transcribe using Whisper.")
     out = pipe(inputs_path, batch_size=BATCH_SIZE, generate_kwargs={"task": task}, return_timestamps=True)
     
     text = out["text"]
     
+    pbar.update(1)
+    pbar.set_description("Merge chunks.")
     chunks = naive_postprocess_whisper_chunks(out["chunks"], inputs, sampling_rate)
 
+    pbar.update(1)
+    pbar.set_description("Create dataset.")
+    
+
     transcripts = []
     audios = []
     with tempfile.TemporaryDirectory() as tmpdirname:
-        for i,chunk in enumerate(chunks):
+        for i,chunk in tqdm.tqdm(enumerate(chunks), desc="Creating dataset (and clean audio if asked for)"):
             
             # TODO: make sure 1D or 2D?
             arr = chunk["audio"]
@@ -79,10 +87,12 @@ def transcribe(inputs_path, task, use_demucs, dataset_name, oauth_token: gr.OAut
             
         dataset = Dataset.from_dict({"audio": audios, "transcript": transcripts}).cast_column("audio", Audio())
         
-    
-        dataset.push_to_hub(dataset_name, token=oauth_token.token)
+        pbar.update(1)
+        pbar.set_description("Push dataset.")
+        dataset.push_to_hub(dataset_name, token=oauth_token.token if oauth_token else oauth_token)
         
-    return  text
+    pbar.close()
+    return text, [[transcript] for transcript in transcripts]
 
 
 def _return_yt_html_embed(yt_url):
@@ -125,11 +135,18 @@ def download_yt_audio(yt_url, filename):
             raise gr.Error(str(err))
 
 
-def yt_transcribe(yt_url, task, use_demucs, dataset_name, oauth_token: gr.OAuthToken | None, max_filesize=75.0, dataset_sampling_rate = 24000):
+def yt_transcribe(yt_url, task, use_demucs, dataset_name, oauth_token: gr.OAuthToken | None, max_filesize=75.0, dataset_sampling_rate = 24000,
+                  progress=gr.Progress(track_tqdm=True)):
+    
+    pbar = tqdm.tqdm(total=5, desc="Overall progression")
+
     html_embed_str = _return_yt_html_embed(yt_url)
 
     with tempfile.TemporaryDirectory() as tmpdirname:
         filepath = os.path.join(tmpdirname, "video.mp4")
+        pbar.update(1)
+        pbar.set_description("Download Youtube video.")
+        
         download_yt_audio(yt_url, filepath)
         with open(filepath, "rb") as f:
             inputs_path = f.read()
@@ -137,18 +154,25 @@ def yt_transcribe(yt_url, task, use_demucs, dataset_name, oauth_token: gr.OAuthT
         inputs = ffmpeg_read(inputs_path, pipe.feature_extractor.sampling_rate)
         inputs = {"array": inputs, "sampling_rate": pipe.feature_extractor.sampling_rate}
     
+        pbar.update(1)
+        pbar.set_description("Transcribe using Whisper.")
         out = pipe(inputs, batch_size=BATCH_SIZE, generate_kwargs={"task": task}, return_timestamps=True)
         
         text = out["text"]
             
         inputs = ffmpeg_read(inputs_path, dataset_sampling_rate)
     
+    pbar.update(1)
+    pbar.set_description("Merge chunks.")
     chunks = naive_postprocess_whisper_chunks(out["chunks"], inputs, dataset_sampling_rate)
 
+    pbar.update(1)
+    pbar.set_description("Create dataset.")
+    
     transcripts = []
     audios = []
     with tempfile.TemporaryDirectory() as tmpdirname:
-        for i,chunk in enumerate(chunks):
+        for i,chunk in tqdm.tqdm(enumerate(chunks), desc="Creating dataset (and clean audio if asked for)."):
             
             # TODO: make sure 1D or 2D?
             arr = chunk["audio"]
@@ -165,23 +189,28 @@ def yt_transcribe(yt_url, task, use_demucs, dataset_name, oauth_token: gr.OAuthT
             
         dataset = Dataset.from_dict({"audio": audios, "transcript": transcripts}).cast_column("audio", Audio())
         
-    
-        dataset.push_to_hub(dataset_name, token=oauth_token.token)
+        pbar.update(1)
+        pbar.set_description("Push dataset.")
+        dataset.push_to_hub(dataset_name, token=oauth_token.token if oauth_token else oauth_token)
         
+    pbar.close()
 
-    return html_embed_str, text
+    return html_embed_str, text, [[transcript] for transcript in transcripts]
 
 
 def naive_postprocess_whisper_chunks(chunks, audio_array, sampling_rate,  stop_chars = ".!:;?", min_duration = 5):
     # merge chunks as long as merged audio duration is lower than min_duration and that a stop character is not met
     # return list of dictionnaries (text, audio)
     # min duration is in seconds
-    
+    pbar = tqdm.tqdm(total=len(chunks), desc="Post-processing transcribed chunks")
     min_duration = int(min_duration * sampling_rate)
+
     
     new_chunks = []
     while chunks:
         current_chunk = chunks.pop(0)
+        pbar.update(1)
+        
         begin, end = current_chunk["timestamp"]
         begin, end = int(begin*sampling_rate), int(end*sampling_rate)
         
@@ -193,7 +222,7 @@ def naive_postprocess_whisper_chunks(chunks, audio_array, sampling_rate,  stop_c
         chunk_to_concat = [audio_array[begin:end]]
         while chunks and (text[-1] not in stop_chars or (current_dur<min_duration)):
             ch = chunks.pop(0)
-            
+            pbar.update(1)
             begin, end = ch["timestamp"]
             begin, end = int(begin*sampling_rate), int(end*sampling_rate)
             current_dur += end-begin
@@ -209,53 +238,75 @@ def naive_postprocess_whisper_chunks(chunks, audio_array, sampling_rate,  stop_c
             "audio": np.concatenate(chunk_to_concat),
         })
         print(f"LENGTH CHUNK #{len(new_chunks)}: {current_dur/sampling_rate}s")
+    
+    pbar.close()
             
     return new_chunks
     
     
-    
-mf_transcribe = gr.Interface(
-    fn=transcribe,
-    inputs=[
-        gr.Audio(type="filepath"),
-        gr.Radio(["transcribe", "translate"], label="Task", value="transcribe"),
-        gr.Radio(["no-post-processing", "separate-audio"], label="Audio separation and cleaning (takes longer - use it if your samples are not cleaned (background noise and music))", value="separate-audio"),
-        gr.Textbox(lines=1, placeholder="Place your new dataset name here", label="Dataset name"),
-    ],
-    outputs="text",
-    theme="huggingface",
-    title="Create your own TTS dataset using your own recordings",
-    description=(
-        "This demo allows use to create a text-to-speech dataset from an input audio snippet and push it to hub to keep track of it."
-        f"Demo uses the checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and 🤗 Transformers to automatically transcribe audio files"
-        " of arbitrary length. It then merge chunks of audio and push it to the hub."
-    ),
-    allow_flagging="never",
-)
-
-yt_transcribe = gr.Interface(
-    fn=yt_transcribe,
-    inputs=[
-        gr.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL"),
-        gr.Radio(["transcribe", "translate"], label="Task", value="transcribe"),
-        gr.Radio(["no-post-processing", "separate-audio"], label="Audio separation and cleaning (takes longer - use it if your samples are not cleaned (background noise and music))", value="separate-audio"),
-        gr.Textbox(lines=1, placeholder="Place your new dataset name here", label="Dataset name"),
-    ],
-    outputs=["html", "text"],
-    theme="huggingface",
-    title="Create your own TTS dataset using Youtube",
-    description=(
-        "This demo allows use to create a text-to-speech dataset from an input audio snippet and push it to hub to keep track of it."
-        f"Demo uses the checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and 🤗 Transformers to automatically transcribe audio files"
-        " of arbitrary length. It then merge chunks of audio and push it to the hub."
-    ),
-    allow_flagging="never",
-)
 
-with gr.Blocks() as demo:
+css = """
+#container{
+    margin: 0 auto;
+    max-width: 80rem;
+}
+#intro{
+    max-width: 100%;
+    text-align: center;
+    margin: 0 auto;
+}
+"""
+with gr.Blocks(css=css) as demo:
     with gr.Row():
         gr.LoginButton().activate()
         gr.LogoutButton()
-    gr.TabbedInterface([mf_transcribe, yt_transcribe], ["Microphone or Audio file", "YouTube"])
-
-demo.launch(debug=True)
+        
+    with gr.Tab("Microphone or Audio file"):
+        gr.Markdown("Create your own TTS dataset using your own recordings", elem_id="intro")
+        gr.Markdown(        
+    "This demo allows use to create a text-to-speech dataset from an input audio snippet and push it to hub to keep track of it."
+    f"Demo uses the checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and 🤗 Transformers to automatically transcribe audio files"
+    " of arbitrary length. It then merge chunks of audio and push it to the hub."
+    )
+        with gr.Column():
+            audio_file = gr.Audio(type="filepath")
+            task_file = gr.Radio(["transcribe", "translate"], label="Task", value="transcribe")
+            cleaning_file = gr.Radio(["no-post-processing", "separate-audio"], label="Audio separation and cleaning (takes longer - use it if your samples are not cleaned (background noise and music))", value="separate-audio")
+            textbox_file = gr.Textbox(lines=1, placeholder="Place your new dataset name here", label="Dataset name")
+            
+            with gr.Row():
+                clear_file = gr.ClearButton([audio_file, task_file, cleaning_file, textbox_file])
+                submit_file = gr.Button("Submit")
+            
+        with gr.Column():
+            transcript_file = gr.Textbox(label="Transcription")
+            dataset_file = gr.Dataset(components=["text"], headers=["Transcripts"])
+            
+    
+    with gr.Tab("YouTube"):
+        gr.Markdown("Create your own TTS dataset using Youtube", elem_id="intro")
+        gr.Markdown(        
+    "This demo allows use to create a text-to-speech dataset from an input audio snippet and push it to hub to keep track of it."
+    f"Demo uses the checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and 🤗 Transformers to automatically transcribe audio files"
+    " of arbitrary length. It then merge chunks of audio and push it to the hub."
+    )
+        with gr.Column():
+            audio_youtube = gr.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL")
+            task_youtube = gr.Radio(["transcribe", "translate"], label="Task", value="transcribe")
+            cleaning_youtube = gr.Radio(["no-post-processing", "separate-audio"], label="Audio separation and cleaning (takes longer - use it if your samples are not cleaned (background noise and music))", value="separate-audio")
+            textbox_youtube = gr.Textbox(lines=1, placeholder="Place your new dataset name here", label="Dataset name")
+            
+            with gr.Row():
+                clear_youtube = gr.ClearButton([audio_youtube, task_youtube, cleaning_youtube, textbox_youtube])
+                submit_youtube = gr.Button("Submit")
+            
+        with gr.Column():
+            html_youtube = gr.HTML()
+            transcript_youtube = gr.Textbox(label="Transcription")
+            dataset_youtube = gr.Dataset(components=["text"], headers=["Transcripts"])
+                
+    
+    submit_file.click(transcribe, inputs=[audio_file, task_file, cleaning_file, textbox_file], outputs=[transcript_file, dataset_file])                
+    submit_youtube.click(yt_transcribe, inputs=[audio_youtube, task_youtube, cleaning_youtube, textbox_youtube], outputs=[html_youtube, transcript_youtube, dataset_youtube])                
+    
+demo.launch(debug=True)