VoiceVista

Video and document translation. Won the Principal's Choice Award at TechXcellence-2025.

Role
Team hackathon project
Period
2025
Stack
Flask, Whisper, Google Translate, gTTS, FFmpeg, Tesseract OCR
Repo
github.com/devnxh/VoiceVista
Exhibit C · VoiceVista

VoiceVista translates the speech in a video into another language and dubs it back over the video. It also extracts and translates text from documents and images. It was built by a team at a hackathon.

What it does #

  • Upload a video, or give a YouTube URL.
  • Transcribe the speech with OpenAI’s Whisper model.
  • Translate the transcript and merge the translated audio with the original video.
  • Summarise the content to about 25% of its original length.
  • Extract and translate text from PDF and Word documents.
  • Read text from images with optical character recognition.
  • Download both the original and the translated text.

It supports 11 languages: English, Hindi, Spanish, French, German, Gujarati, Urdu, Bengali, Tamil, Marathi and Kannada.

How video translation works #

flowchart LR
  A[Video or YouTube URL] --> B[FFmpeg extracts audio]
  B --> C[Whisper transcribes]
  C --> D[Summary]
  D --> E[Google Translate]
  E --> F[gTTS speaks the translation]
  F --> G[FFmpeg merges audio and video]

How document processing works #

Input How the text is extracted
PDF PyPDF2 reads every page
Word (.docx) python-docx reads every paragraph
Image The Tesseract OCR engine reads the text

The extracted text is cleaned, translated with Google Translate, and shown next to the original.

Limits #

  • Long videos take time, mostly in the Whisper transcription, and need more memory.
  • OCR quality depends on how clear the image is.
  • PDF extraction works best on digitally created PDFs, not scans.

Award #

VoiceVista won the Principal’s Choice Award at TechXcellence-2025, held on 4 and 5 April 2025 at Government Engineering College Bhavnagar and organised by B2B and the GDG GECG Bhavnagar Chapter.

What I would change #

The weakest link is the translation and voice step. Google Translate and gTTS get the words across but lose tone and timing, so the dubbed audio sounds flat and rarely lines up with the speaker. Given more time I would move to models that keep prosody and align to the original speech, and cache Whisper transcriptions so a long video does not pay the transcription cost twice.

Contents