SDT — Subtitle from Transcript Generator

Generate accurate .srt subtitle files by aligning a plain text transcript (no timestamps) to a YouTube video or local audio/video file.

How It Works

SDT uses forced alignment — it takes your known transcript and synchronizes it with the audio using stable-ts (built on OpenAI Whisper). This is much faster and more accurate than transcribing from scratch, because the model already knows what is being said and only needs to figure out when.

Input: YouTube URL + Plain Transcript  →  Output: Timed .srt Subtitle File

Features

• 🎬 YouTube support — paste a URL, audio is downloaded automatically via yt-dlp
• 📁 Local files — works with any video/audio format (MP4, MKV, MP3, WAV, etc.)
• 🌍 Multilingual — supports all Whisper languages (English, Chinese, Japanese, etc.)
• 🚀 GPU acceleration — auto-detects CUDA GPU for fast alignment
• ✂️ Smart segmentation — automatically splits subtitles at natural breakpoints
• 📓 Colab ready — included notebook for easy cloud usage with free GPU

Quick Start

Installation

pip install -r requirements.txt

Note: FFmpeg must be installed on your system for YouTube downloads.

CLI Usage

# From YouTube
python -m sdt -i "https://youtube.com/watch?v=VIDEO_ID" -t transcript.txt -o output.srt

# From local file
python -m sdt -i video.mp4 -t transcript.txt

# Chinese with large model
python -m sdt -i video.mp4 -t transcript.txt -l zh -m large-v3

# Preview without saving
python -m sdt -i audio.mp3 -t script.txt --preview

Python API

from sdt import download_audio, align_transcript
from sdt.srt_writer import write_srt

# 1. Get audio
audio_path = download_audio("https://youtube.com/watch?v=VIDEO_ID")

# 2. Align transcript
with open("transcript.txt", "r") as f:
    transcript = f.read()

result = align_transcript(audio_path, transcript, language="en")

# 3. Generate SRT
write_srt(result, "output.srt")

Google Colab

Open SDT_Colab.ipynb in Google Colab for a ready-to-use notebook with free GPU.

CLI Options

Flag	Description	Default
-i, --input	YouTube URL or local file path	required
-t, --transcript	Path to plain text transcript	required
-o, --output	Output file path	auto-named
-l, --language	Language code (en, zh, ja, ...)	auto-detect
-m, --model	Whisper model size	medium
--max-chars	Max characters per subtitle	42
--max-duration	Max seconds per subtitle	5.0
--format	Output format (srt or vtt)	srt
--preview	Print to console, don't save	off

Model Sizes

Model	Parameters	English	Multilingual	Speed
tiny	39M	⭐⭐	⭐	⚡⚡⚡⚡
base	74M	⭐⭐⭐	⭐⭐	⚡⚡⚡
small	244M	⭐⭐⭐⭐	⭐⭐⭐	⚡⚡
medium	769M	⭐⭐⭐⭐⭐	⭐⭐⭐⭐	⚡
large-v3	1.5B	⭐⭐⭐⭐⭐	⭐⭐⭐⭐⭐	🐢

Requirements

• Python ≥ 3.9
• FFmpeg (for YouTube downloads)
• NVIDIA GPU (optional, but recommended for speed)