app.py

import torch
import gradio as gr
import yt_dlp as youtube_dl
from transformers import pipeline
from transformers.pipelines.audio_utils import ffmpeg_read
from urllib.parse import urlparse, parse_qs

import tempfile
import time
import os
import numpy as np

# Constants
MODEL_NAME = "dataprizma/whisper-large-v3-turbo"
BATCH_SIZE = 8
FILE_LIMIT_MB = 1000
YT_LENGTH_LIMIT_S = 3600  # 1 hour limit

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

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

# Extract YouTube Video ID
def _extract_yt_video_id(yt_url):
    parsed_url = urlparse(yt_url)
    return parse_qs(parsed_url.query).get("v", [""])[0]

# Embed YouTube Video in HTML
def _return_yt_html_embed(yt_url):
    video_id = _extract_yt_video_id(yt_url)
    if not video_id:
        raise gr.Error("Invalid YouTube URL. Please check and try again.")
    return f'<center> <iframe width="500" height="320" src="https://www.youtube.com/embed/{video_id}"></iframe> </center>'

# Transcription function (Fix applied)
def transcribe(audio_file, task):
    if audio_file is None:
        raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting.")

    # Open file as binary to ensure correct data type
    with open(audio_file, "rb") as f:
        audio_data = f.read()

    # Read audio using ffmpeg_read (correcting input format)
    audio_array = ffmpeg_read(audio_data, pipe.feature_extractor.sampling_rate)

    # Convert to proper format
    inputs = {
        "raw": np.array(audio_array),  
        "sampling_rate": pipe.feature_extractor.sampling_rate
    }

    # Perform transcription
    result = pipe(
        inputs,
        batch_size=BATCH_SIZE,
        generate_kwargs={"task": task},
        return_timestamps=True
    )

    return result["text"]
# Download YouTube audio
def download_yt_audio(yt_url, filename):
    ydl_opts = {
        "format": "bestaudio/best",
        "outtmpl": filename,
        "postprocessors": [
            {"key": "FFmpegExtractAudio", "preferredcodec": "mp3", "preferredquality": "192"}
        ],
    }
    
    with youtube_dl.YoutubeDL(ydl_opts) as ydl:
        try:
            info = ydl.extract_info(yt_url, download=False)
            file_length_s = info.get("duration", 0)  # Duration in seconds
            if file_length_s > YT_LENGTH_LIMIT_S:
                raise gr.Error(f"Maximum YouTube length is 1 hour. Your video is {file_length_s // 3600}h {file_length_s % 3600 // 60}m {file_length_s % 60}s.")
            ydl.download([yt_url])
        except youtube_dl.utils.DownloadError as err:
            raise gr.Error(str(err))

# YouTube transcription function
def yt_transcribe(yt_url, task, max_filesize=75.0):
    html_embed_str = _return_yt_html_embed(yt_url)

    with tempfile.TemporaryDirectory() as tmpdirname:
        filepath = os.path.join(tmpdirname, "audio.mp3")
        download_yt_audio(yt_url, filepath)

        if os.path.getsize(filepath) > max_filesize * 1024 * 1024:
            raise gr.Error(f"File too large! Max allowed size is {max_filesize}MB.")

        with open(filepath, "rb") as f:
            inputs = ffmpeg_read(f.read(), pipe.feature_extractor.sampling_rate)

    inputs = {
        "array": inputs,
        "sampling_rate": pipe.feature_extractor.sampling_rate,
        "attention_mask": torch.ones(len(inputs), dtype=torch.long),
    }

    text = pipe(
        {"input_features": inputs}, 
        batch_size=BATCH_SIZE, 
        generate_kwargs={"task": task, "forced_decoder_ids": None}, 
        return_timestamps=True
    )["text"]

    return html_embed_str, text

# Gradio UI
demo = gr.Blocks()

file_transcribe = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.Audio(type="filepath", label="Audio file"),
        gr.Radio(["transcribe", "translate"], label="Task"),
    ],
    outputs="text",
    title="Whisper Large V3: Transcribe Audio",
    description="Whisper Large V3 fine-tuned for Uzbek language by Dataprizma",
    flagging_mode="never",
)

yt_transcribe = gr.Interface(
    fn=yt_transcribe,
    inputs=[
        gr.Textbox(lines=1, placeholder="Paste YouTube URL here", label="YouTube URL"),
        gr.Radio(["transcribe", "translate"], label="Task")
    ],
    outputs=["html", "text"],
    title="Whisper Large V3: Transcribe YouTube",
    description="Whisper Large V3 fine-tuned for Uzbek language by Dataprizma",
    flagging_mode="never",
)

with demo:
    gr.TabbedInterface([file_transcribe, yt_transcribe], ["Audio file", "YouTube"])

demo.launch()