app.py

import os
import math
import gradio as gr
import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
from moviepy.editor import AudioFileClip

def transcribe(audio_file, transcribe_to_text, transcribe_to_srt, language):
    device = "cuda:0" if torch.cuda.is_available() else "cpu"
    torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

    model_id = "openai/whisper-large-v3"
    model = AutoModelForSpeechSeq2Seq.from_pretrained(
        model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
    )
    model.to(device)

    processor = AutoProcessor.from_pretrained(model_id)

    pipe = pipeline(
        "automatic-speech-recognition",
        model=model,
        tokenizer=processor.tokenizer,
        feature_extractor=processor.feature_extractor,
        max_new_tokens=128,
        chunk_length_s=30,
        batch_size=2,
        return_timestamps=True,
        torch_dtype=torch_dtype,
        device=device,
        generate_kwargs={"language": language}
    )

    audio = AudioFileClip(audio_file.name)
    duration = audio.duration
    n_chunks = math.ceil(duration / 30)

    transcription_txt = ""
    transcription_srt = []

    for i in range(n_chunks):
        start = i * 30
        end = min((i + 1) * 30, duration)
        audio_chunk = audio.subclip(start, end)
        
        temp_file_path = f"temp_audio_{i}.wav"
        audio_chunk.write_audiofile(temp_file_path, codec='pcm_s16le')

        with open(temp_file_path, "rb") as temp_file:
            result = pipe(temp_file_path)
            transcription_txt += result["text"]

            if transcribe_to_srt:
                for chunk in result["chunks"]:
                    start_time, end_time = chunk["timestamp"]
                    transcription_srt.append({
                        "start": start_time + i * 30,
                        "end": end_time + i * 30,
                        "text": chunk["text"]
                    })

        os.remove(temp_file_path)

        yield f"Progress: {int(((i + 1) / n_chunks) * 100)}%"

    output = ""
    if transcribe_to_text:
        output += "Text Transcription:\n" + transcription_txt + "\n\n"

    if transcribe_to_srt:
        output += "SRT Transcription:\n"
        for i, sub in enumerate(transcription_srt, 1):
            output += f"{i}\n{format_time(sub['start'])} --> {format_time(sub['end'])}\n{sub['text']}\n\n"

    yield output

def format_time(seconds):
    m, s = divmod(seconds, 60)
    h, m = divmod(m, 60)
    return f"{int(h):02d}:{int(m):02d}:{s:06.3f}".replace('.', ',')

iface = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.Audio(type="filepath"),
        gr.Checkbox(label="Transcribe to Text"),
        gr.Checkbox(label="Transcribe to SRT"),
        gr.Dropdown(choices=['en', 'he', 'it', 'fr', 'de', 'zh', 'ar'], label="Language")
    ],
    outputs="text",
    title="WhisperCap Transcription",
    description="Upload an audio file to transcribe it using Whisper.",
)

iface.launch()