Real time speech transcription, English to Hindi translation, and handwriting recognition, wrapped in one Django REST API.
Built for the Intel IEEE INDICON 2024 track, 30 November to 4 December 2024.
India has over 1,600 languages. A student who grew up speaking one of them can end up in a classroom, seminar or workshop delivered in another, and the barrier is not intelligence but medium. The gap shows up in lectures they cannot follow and notes they cannot read back.
So the tool does three things at the point of use: hear it, read it, and turn it into the language the student actually thinks in.
Audio goes to Google Cloud Speech-to-Text at 16 kHz LINEAR16, and the transcript
comes back into the page. From there it can be translated to Hindi through a
local MarianMT model, Helsinki-NLP/opus-mt-en-hi, running through Hugging Face
transformers rather than a paid translation API.
The translation is deliberately narrow. translate_text() raises if the target
language is anything other than Hindi, rather than silently returning bad output
for a pair the model was never trained on.
Handwritten Devanagari characters are read off an image and returned as text, so a textbook page or a set of handwritten study notes becomes something that can be searched and translated.
This is the part that runs on Intel OpenVINO. The model is loaded as IR
through openvino.runtime.Core, compiled for CPU, and called directly. That is
the reason the project fits the Intel track: inference stays local on ordinary
CPU hardware rather than requiring a GPU or a hosted endpoint.
Django REST Framework, JWT auth through rest_framework_simplejwt.
| Method | Route | Auth | Purpose |
|---|---|---|---|
| POST | /api/signup/ |
no | Create a user |
| POST | /api/login/ |
no | Returns access and refresh tokens |
| POST | /api/ai-processing/ |
yes | OpenVINO inference on submitted input |
| POST | /api/speech-to-text/ |
yes | Upload audio, get a transcript |
| POST | /api/translate/ |
yes | English to Hindi |
The three AI routes all require a valid token. Uploaded audio is written to
./uploads/, transcribed, then deleted in the same request.
pip install django djangorestframework djangorestframework-simplejwt transformers openvino google-cloud-speech python-dotenv torchCopy .env.example to .env and fill in both paths, then:
python manage.py migrate && python manage.py runserverMODEL_PATH must point at the OpenVINO IR .xml file.
GOOGLE_APPLICATION_CREDENTIALS must point at a Google Cloud service account
JSON with the Speech-to-Text API enabled. Neither file is in this repository.
Published as written during the competition, not tidied up afterwards.
views.py and backend.py are near duplicates. Both define the same five
API classes and the same three helper functions. urls.py imports from
views.py, so backend.py is the dead copy. It differs only in calling
django.setup() and building a WSGI application at import time, which is what
made it a scratch entry point during development.
There is a second, unrelated Flask app. app.py serves templates/ with its
own login and signup, backed by a Python dict holding one demo account. It shares
no code and no user store with the Django API. It was the front end shell while
the API was being built.
models.py and serializers.py are empty. The project uses Django's built-in
User and hand-rolled request parsing instead.
DEBUG = True and the Django SECRET_KEY is a placeholder rather than a
generated key. Both would need changing before this was exposed to anything.
The default MODEL_PATH in backend.py points at a directory, not a model
file. It falls back to the virtual environment folder, which cannot be read as
IR, so compiled_model ends up None and every inference call raises. The
.env path is what actually made it work.
This repo was 118 MB across 33,878 files for a project whose source is 17 files.
venv/ and openvino_env/ were both committed, so every clone pulled two
complete Python environments including a 42 MB OpenVINO CPU plugin binary and a
38 MB OpenBLAS DLL.
History has been rewritten to remove them, along with db.sqlite, the committed
.env, and a stray tempCodeRunnerFile.py. The repository is now 404 KB.
Commit dates and authorship are unchanged.
.env.example replaces the committed .env. Neither ever held a secret, only
local paths, but the real one does not belong in git.
Aditya Bhaty · Aditya Naidu
Commits from this period appear under Adi-ctive and Faith1406.
| Demo video | https://www.youtube.com/watch?v=SXIICG16GOc |
The two clips above are from the demo section of that video. The rest of it is stock footage, plus third-party application recordings that belong to their respective owners.