app.py

import torch

import gradio as gr
import yt_dlp as youtube_dl
import numpy as np
from datasets import Dataset, Audio
from scipy.io import wavfile

from transformers import pipeline
from transformers.pipelines.audio_utils import ffmpeg_read

import tempfile
import os
import time
import demucs.api



MODEL_NAME = "openai/whisper-large-v3"
DEMUCS_MODEL_NAME = "htdemucs_ft"
BATCH_SIZE = 8
FILE_LIMIT_MB = 1000
YT_LENGTH_LIMIT_S = 3600  # limit to 1 hour YouTube files

device = 0 if torch.cuda.is_available() else "cpu"

pipe = pipeline(
    task="automatic-speech-recognition",
    model=MODEL_NAME,
    chunk_length_s=30,
    device=device,
)

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 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):
    if inputs_path is None:
        raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting your request.")
    
    sampling_rate, inputs = wavfile.read(inputs_path) 

    out = pipe(inputs_path, batch_size=BATCH_SIZE, generate_kwargs={"task": task}, return_timestamps=True)
    
    text = out["text"]
    
    chunks = naive_postprocess_whisper_chunks(out["chunks"], inputs, sampling_rate)

    transcripts = []
    audios = []
    with tempfile.TemporaryDirectory() as tmpdirname:
        for i,chunk in enumerate(chunks):
            
            # TODO: make sure 1D or 2D?
            arr = chunk["audio"]
            path = os.path.join(tmpdirname, f"{i}.wav")
            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, "transcript": transcripts}).cast_column("audio", Audio())
        
    
        dataset.push_to_hub(dataset_name, token=oauth_token)
        
    return  text


def _return_yt_html_embed(yt_url):
    video_id = yt_url.split("?v=")[-1]
    HTML_str = (
        f'<center> <iframe width="500" height="320" src="https://www.youtube.com/embed/{video_id}"> </iframe>'
        " </center>"
    )
    return HTML_str

def download_yt_audio(yt_url, filename):
    info_loader = youtube_dl.YoutubeDL()
    
    try:
        info = info_loader.extract_info(yt_url, download=False)
    except youtube_dl.utils.DownloadError as err:
        raise gr.Error(str(err))
    
    file_length = info["duration_string"]
    file_h_m_s = file_length.split(":")
    file_h_m_s = [int(sub_length) for sub_length in file_h_m_s]
    
    if len(file_h_m_s) == 1:
        file_h_m_s.insert(0, 0)
    if len(file_h_m_s) == 2:
        file_h_m_s.insert(0, 0)
    file_length_s = file_h_m_s[0] * 3600 + file_h_m_s[1] * 60 + file_h_m_s[2]
    
    if file_length_s > YT_LENGTH_LIMIT_S:
        yt_length_limit_hms = time.strftime("%HH:%MM:%SS", time.gmtime(YT_LENGTH_LIMIT_S))
        file_length_hms = time.strftime("%HH:%MM:%SS", time.gmtime(file_length_s))
        raise gr.Error(f"Maximum YouTube length is {yt_length_limit_hms}, got {file_length_hms} YouTube video.")
    
    ydl_opts = {"outtmpl": filename, "format": "worstvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best"}
    
    with youtube_dl.YoutubeDL(ydl_opts) as ydl:
        try:
            ydl.download([yt_url])
        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, max_filesize=75.0, dataset_sampling_rate = 24000):
    html_embed_str = _return_yt_html_embed(yt_url)

    with tempfile.TemporaryDirectory() as tmpdirname:
        filepath = os.path.join(tmpdirname, "video.mp4")
        download_yt_audio(yt_url, filepath)
        with open(filepath, "rb") as f:
            inputs_path = f.read()

    inputs = ffmpeg_read(inputs_path, pipe.feature_extractor.sampling_rate)
    inputs = {"array": inputs, "sampling_rate": pipe.feature_extractor.sampling_rate}

    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)
    
    chunks = naive_postprocess_whisper_chunks(out["chunks"], inputs, dataset_sampling_rate)

    transcripts = []
    audios = []
    with tempfile.TemporaryDirectory() as tmpdirname:
        for i,chunk in enumerate(chunks):
            
            # TODO: make sure 1D or 2D?
            arr = chunk["audio"]
            path = os.path.join(tmpdirname, f"{i}.wav")
            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, "transcript": transcripts}).cast_column("audio", Audio())
        
    
        dataset.push_to_hub(dataset_name, token=oauth_token)
        

    return html_embed_str, text


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
        
        text = current_chunk["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
    
    
    
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:
    with gr.Row():
        gr.LoginButton()
        gr.LogoutButton()
    gr.TabbedInterface([mf_transcribe, yt_transcribe], ["Microphone or Audio file", "YouTube"])

demo.launch(debug=True)