Update app.py

app.py

@@ -1,4 +1,46 @@
+Hugging Face's logo
+Hugging Face
+Search models, datasets, users...
+Models
+Datasets
+Spaces
+Posts
+Docs
+Pricing
+
+
+
+Spaces:
+
+ylacombe
+/
+create-your-own-TTS-dataset
+
+
+like
+21
+App
+Files
+Community
+2
+create-your-own-TTS-dataset
+/
+app.py
+
+ylacombe's picture
+ylacombe
+HF STAFF
+Update app.py
+35d4f1f
+7 months ago
+raw
+history
+blame
+No virus
+
+13.5 kB
 import torch
+
 import gradio as gr
 import yt_dlp as youtube_dl
 import numpy as np
@@ -11,9 +53,9 @@ from transformers.pipelines.audio_utils import ffmpeg_read
 import tempfile
 import os
 import time
-import demucs
+import demucs.api
 
-MODEL_NAME = "openai/whisper-large-v2"
+MODEL_NAME = "openai/whisper-large-v3"
 DEMUCS_MODEL_NAME = "htdemucs_ft"
 BATCH_SIZE = 8
 FILE_LIMIT_MB = 1000
@@ -28,13 +70,14 @@ pipe = pipeline(
     device=device,
 )
 
-separator = demucs.api.Separator(model=DEMUCS_MODEL_NAME)
+separator = demucs.api.Separator(model = DEMUCS_MODEL_NAME, )
 
 def separate_vocal(path):
     origin, separated = separator.separate_audio_file(path)
     demucs.api.save_audio(separated["vocals"], path, samplerate=separator.samplerate)
     return path
 
+    
 def transcribe(inputs_path, task, use_demucs, dataset_name, oauth_token: gr.OAuthToken | None, progress=gr.Progress()):
     if inputs_path is None:
         raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting your request.")
@@ -51,7 +94,7 @@ def transcribe(inputs_path, task, use_demucs, dataset_name, oauth_token: gr.OAut
     current_step += 1
     progress((current_step, total_step), desc="Transcribe using Whisper.")
 
-    sampling_rate, inputs = wavfile.read(inputs_path)
+    sampling_rate, inputs = wavfile.read(inputs_path) 
 
     out = pipe(inputs_path, batch_size=BATCH_SIZE, generate_kwargs={"task": task}, return_timestamps=True)
     
@@ -64,21 +107,25 @@ def transcribe(inputs_path, task, use_demucs, dataset_name, oauth_token: gr.OAut
     current_step += 1
     progress((current_step, total_step), desc="Create dataset.")
     
+
     transcripts = []
     audios = []
     with tempfile.TemporaryDirectory() as tmpdirname:
-        for i, chunk in enumerate(progress.tqdm(chunks, desc="Creating dataset (and clean audio if asked for)")):
+        for i,chunk in enumerate(progress.tqdm(chunks, desc="Creating dataset (and clean audio if asked for)")):
+            
+            # TODO: make sure 1D or 2D?
             arr = chunk["audio"]
             path = os.path.join(tmpdirname, f"{i}.wav")
-            wavfile.write(path, sampling_rate, arr)
+            wavfile.write(path, sampling_rate,  arr)
             
             if use_demucs == "separate-audio":
+                # use demucs tp separate vocals
                 print(f"Separating vocals #{i}")
                 path = separate_vocal(path)
                 
             audios.append(path)
             transcripts.append(chunk["text"])
-        
+            
         dataset = Dataset.from_dict({"audio": audios, "text": transcripts}).cast_column("audio", Audio())
         
         current_step += 1
@@ -87,6 +134,7 @@ def transcribe(inputs_path, task, use_demucs, dataset_name, oauth_token: gr.OAut
         
     return [[transcript] for transcript in transcripts], text
 
+
 def _return_yt_html_embed(yt_url):
     video_id = yt_url.split("?v=")[-1]
     HTML_str = (
@@ -126,7 +174,8 @@ def download_yt_audio(yt_url, filename):
         except youtube_dl.utils.ExtractorError as err:
             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()):
     
     if yt_url is None:
@@ -141,7 +190,7 @@ def yt_transcribe(yt_url, task, use_demucs, dataset_name, oauth_token: gr.OAuthT
 
     if oauth_token is None:
         gr.Warning("Make sure to click and login before using this demo.")
-        return html_embed_str, [["transcripts will appear here"]], ""
+        return  html_embed_str, [["transcripts will appear here"]], ""
         
     current_step += 1
     progress((current_step, total_step), desc="Load video.")
@@ -174,102 +223,133 @@ def yt_transcribe(yt_url, task, use_demucs, dataset_name, oauth_token: gr.OAuthT
     transcripts = []
     audios = []
     with tempfile.TemporaryDirectory() as tmpdirname:
-        for i, chunk in enumerate(progress.tqdm(chunks, desc="Creating dataset (and clean audio if asked for).")):
+        for i,chunk in enumerate(progress.tqdm(chunks, desc="Creating dataset (and clean audio if asked for).")):
+            
+            # TODO: make sure 1D or 2D?
             arr = chunk["audio"]
             path = os.path.join(tmpdirname, f"{i}.wav")
-            wavfile.write(path, dataset_sampling_rate, arr)
+            wavfile.write(path, dataset_sampling_rate,  arr)
             
             if use_demucs == "separate-audio":
+                # use demucs tp separate vocals
                 print(f"Separating vocals #{i}")
                 path = separate_vocal(path)
                 
             audios.append(path)
             transcripts.append(chunk["text"])
-        
+            
         dataset = Dataset.from_dict({"audio": audios, "text": transcripts}).cast_column("audio", Audio())
         
         current_step += 1
         progress((current_step, total_step), desc="Push dataset.")
         dataset.push_to_hub(dataset_name, token=oauth_token.token if oauth_token else oauth_token)
+        
 
     return html_embed_str, [[transcript] for transcript in transcripts], text
 
-def naive_postprocess_whisper_chunks(chunks, audio_array, sampling_rate, stop_chars=".!:;?", min_duration=5):
+
+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
     min_duration = int(min_duration * sampling_rate)
 
+    
     new_chunks = []
     while chunks:
         current_chunk = chunks.pop(0)
+        
         begin, end = current_chunk["timestamp"]
-        begin, end = int(begin * sampling_rate), int(end * sampling_rate)
-        current_dur = end - begin
+        begin, end = int(begin*sampling_rate), int(end*sampling_rate)
+        
+        current_dur = end-begin
+        
         text = current_chunk["text"]
         
-        chunk_to_concat = []
-        while chunks and (current_dur < min_duration or text[-1] not in stop_chars):
-            next_chunk = chunks.pop(0)
-            next_text = next_chunk["text"].strip()
-            next_begin, next_end = next_chunk["timestamp"]
-            next_begin, next_end = int(next_begin * sampling_rate), int(next_end * sampling_rate)
-            current_dur += next_end - next_begin
-            text += f" {next_text}"
-            end = next_end
-
-        new_chunks.append(
-            {
-                "audio": np.array(audio_array[begin:end]).astype(np.float32),
-                "text": text,
-            }
-        )
+            
+        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)
+            begin, end = ch["timestamp"]
+            begin, end = int(begin*sampling_rate), int(end*sampling_rate)
+            current_dur += end-begin
+            
+            text = "".join([text, ch["text"]])
+            
+            # TODO: add silence ?
+            chunk_to_concat.append(audio_array[begin:end])
+            
 
+        new_chunks.append({
+            "text": text.strip(),
+            "audio": np.concatenate(chunk_to_concat),
+        })
+        print(f"LENGTH CHUNK #{len(new_chunks)}: {current_dur/sampling_rate}s")
+                
     return new_chunks
-
-with gr.Blocks() as demo:
-    with gr.Row():
-        with gr.Column():
-            gr.Markdown("### Audio or YouTube Video Transcription")
-            with gr.Row():
-                yt_textbox = gr.Textbox(label="YouTube link")
-                yt_button = gr.Button("Transcribe YouTube video")
-        with gr.Column():
-            gr.Markdown("### Upload or Record Audio")
-            local_audio_input = gr.Audio(type="filepath", label="Upload Audio")
-            local_button = gr.Button("Transcribe Local Audio")
-
-    task = gr.Radio(
-        ["transcribe", "translate"],
-        label="Task",
-        value="transcribe",
-    )
-
-    demucs_checkbox = gr.CheckboxGroup(["separate-audio"], label="Apply Demucs (Separate Vocal from Audio)")
-    dataset_name = gr.Textbox(label="Dataset name", placeholder="Dataset name to push to Hugging Face Hub")
     
-    with gr.Row():
-        login_button = gr.Button("Login")
-        login_output = gr.Markdown()
     
-    with gr.Row():
-        output_transcriptions = gr.Dataframe(headers=["Transcriptions"])
-        output_text = gr.Markdown()
-
-    login_button.click(
-        fn=None,
-        inputs=None,
-        outputs=login_output,
-        _js="function() { return window.location = 'https://huggingface.co/login'; }",
-    )
 
-    yt_button.click(
-        yt_transcribe,
-        inputs=[yt_textbox, task, demucs_checkbox, dataset_name, login_button],
-        outputs=[login_output, output_transcriptions, output_text],
+css = """
+#intro{
+    max-width: 100%;
+    text-align: center;
+    margin: 0 auto;
+}
+"""
+with gr.Blocks(css=css) as demo:
+    with gr.Row():
+        gr.LoginButton()
+        gr.LogoutButton()
+        
+    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."
     )
-
-    local_button.click(
-        transcribe,
-        inputs=[local_audio_input, task, demucs_checkbox, dataset_name, login_button],
-        outputs=[output_transcriptions, output_text],
+        with gr.Row():
+            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. Should be in the format : <user>/<dataset_name> or <org>/<dataset_name>. Also accepts <dataset_name>, which will default to the namespace of the logged-in user.", 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()
+                dataset_youtube = gr.Dataset(label="Transcribed samples.",components=["text"], headers=["Transcripts"], samples=[["transcripts will appear here"]])
+                transcript_youtube = gr.Textbox(label="Transcription")
+        
+    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."
     )
-
-demo.launch()
+        with gr.Row():
+            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. Should be in the format : <user>/<dataset_name> or <org>/<dataset_name>. Also accepts <dataset_name>, which will default to the namespace of the logged-in user.", 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():
+                dataset_file = gr.Dataset(label="Transcribed samples.", components=["text"], headers=["Transcripts"], samples=[["transcripts will appear here"]])
+                transcript_file = gr.Textbox(label="Transcription")
+                
+                
+    
+    submit_file.click(transcribe, inputs=[audio_file, task_file, cleaning_file, textbox_file], outputs=[dataset_file, transcript_file])                
+    submit_youtube.click(yt_transcribe, inputs=[audio_youtube, task_youtube, cleaning_youtube, textbox_youtube], outputs=[html_youtube, dataset_youtube, transcript_youtube])                
+    
+demo.launch(debug=True)