pipeline.py

import openai
from pytube import YouTube
import argparse
import os
import io
import whisper
import ffmpeg

parser = argparse.ArgumentParser()
parser.add_argument("--link", help="youtube video link here", default=None, type=str, required=False)
parser.add_argument("--local_path", help="local video path here", default=None, type=str, required=False)
parser.add_argument("--srt_file", help="srt file input path here", default=None, type=str, required=False)  # New argument
parser.add_argument("--download", help="download path", default='./downloads', type=str, required=False)
parser.add_argument("--output_dir", help="translate result path", default='./results', type=str, required=False)
parser.add_argument("--video_name", help="video name", default='placeholder', type=str, required=False)
parser.add_argument("--model_name", help="model name only support text-davinci-003 and gpt-3.5-turbo", type=str, required=False, default="gpt-3.5-turbo")
args = parser.parse_args()

# input should be either video file or youtube video link.
if args.link is None and args.local_path is None and args.srt_file is None:
    print("need video source or srt file")
    exit()

openai.api_key = os.getenv("OPENAI_API_KEY")
DOWNLOAD_PATH = args.download
RESULT_PATH = args.output_dir
VIDEO_NAME = args.video_name
n_threshold = 1000  # Token limit for the GPT-3.5 model
# model_name = "text-davinci-003" # replace this to our own fintune model
model_name = args.model_name
# model_name = "davinci"

# get source audio
if args.link is not None and args.local_path is None:
    # Download audio from YouTube
    video_link = args.link
    try:
        video = YouTube(video_link)
        audio = video.streams.filter(only_audio=True, file_extension='mp4').first()
        audio.download(DOWNLOAD_PATH)
        print('Download Completed!')
    except Exception as e:
        print("Connection Error")
        print(e) 
    audio_path = '{}/{}'.format(DOWNLOAD_PATH, audio.default_filename)
    audio_file = open(audio_path, "rb")
    VIDEO_NAME = audio.default_filename.split('.')[0]
elif args.local_path is not None:
    # Read from local
    audio_file= open(args.local_path, "rb")
    audio_path = args.local_path


# Instead of using the script_en variable directly, we'll use script_input
srt_file_en = args.srt_file 
if srt_file_en is not None: 
    with open(srt_file_en, 'r') as f:
        script_input = f.read()
else:
    # using whisper to perform speech-to-text and save it in <video name>_en.txt under RESULT PATH.
    srt_file_en = "{}/{}_en.srt".format(RESULT_PATH, VIDEO_NAME)
    if not os.path.exists(srt_file_en):
        # use OpenAI API for transcribe
        # transcript = openai.Audio.transcribe("whisper-1", audio_file) 

        # use local whisper model 
        model = whisper.load_model("base") # using base model in local machine (may use large model on our server)
        transcript = model.transcribe(audio_path)

        #Write SRT file
        from whisper.utils import WriteSRT

        with open(srt_file_en, 'w', encoding="utf-8") as srt:
            writer = WriteSRT(RESULT_PATH)
            writer.write_result(transcript, srt)

    # split the video script(open ai prompt limit: about 5000)
    with open(srt_file_en, 'r') as f:
        script_en = f.read()
        script_input = script_en

from srt2ass import srt2ass
assSub_en = srt2ass(srt_file_en, "default", "No", "Modest")
print('ASS subtitle saved as: ' + assSub_en)

# Split the video script by sentences and create chunks within the token limit
n_threshold = 4096  # Token limit for the GPT-3 model
script_split = script_input.split('.')

script_arr = []
script = ""
for sentence in script_split:
    if len(script) + len(sentence) + 1 <= n_threshold:
        script += sentence + '.'
    else:
        script_arr.append(script.strip())
        script = sentence + '.'
if script.strip():
    script_arr.append(script.strip())

# Translate and save
for s in script_arr:
    # using chatgpt model
    if model_name == "gpt-3.5-turbo":
        # print(s + "\n")
        response = openai.ChatCompletion.create(
            model=model_name,
            messages = [
                {"role": "system", "content": "You are a helpful assistant that translates English to Chinese and have decent background in starcraft2."},
                {"role": "user", "content": 'Translate the following English text to Chinese: "{}"'.format(s)}
            ],
            temperature=0.1
        )
        with open(f"{RESULT_PATH}/{VIDEO_NAME}_zh.srt", 'a+') as f:
            f.write(response['choices'][0]['message']['content'].strip())

    if model_name == "text-davinci-003":
        prompt = f"Please help me translate this into Chinese:\n\n{s}\n\n"
        # print(prompt)
        response = openai.Completion.create(
            model=model_name,
            prompt=prompt,
            temperature=0.1,
            max_tokens=2000,
            top_p=1.0,
            frequency_penalty=0.0,
            presence_penalty=0.0
        )

        with open(f"{RESULT_PATH}/{VIDEO_NAME}_zh.srt", 'a+') as f:
            f.write(response['choices'][0]['text'].strip())

assSub_zh = srt2ass(f"{RESULT_PATH}/{VIDEO_NAME}_zh.srt", "default", "No", "Modest")
print('ASS subtitle saved as: ' + assSub_zh)