Releasing Vegam Whisper Family of Models and demoing Malayalam Speech to Text

Talk page for Demo Presentation session presented @ Summit 23, IIIT Kottayam

Talk Page

Kurian Benoy


Sunday, June 11, 2023

Vegam Whisper Family of Models

Inspired by faster-whisper which is a reimplementation of OpenAI’s Whisper model using CTranslate2 speeds upto 4 times faster than openai/whisper for the same accuracy while using less memory.

CTranslate2 supports various quantization formats and we have trained models for the following quantization formats:

  • float16
  • int16
  • float8
  • int8_float8
  • No quantization

For Malayalam, I used the best model in my benchmarking study ie thennal/whisper-medium-ml and converted into all possible quantization weights.

  1. kurianbenoy/vegam-whisper-medium-ml
  2. kurianbenoy/vegam-whisper-medium-ml-fp16
  3. kurianbenoy/vegam-whisper-medium-ml-int16
  4. kurianbenoy/vegam-whisper-medium-ml-fp8
  5. kurianbenoy/vegam-whisper-medium-ml-int8_float8

Code to run faster-whisper code

Demo Video -1 and it’s associated Output

Oru Thai Nadam sang by Venugopal and Sreya, Lyrics by Sugathakumari

Output of clip from Video 1

Demo Video -2 and it’s associated Output

Sang by Sithara Krishna Kumar, Lyrics by BK Hari Narayanan. This was a song created spontaneously at MBIFL 2023

Output of clip from Video 2


  • Pallakku is a Malayalam speech to text demo leveraging the model-weights of whisper family of models. It might be the first huggingface 🤗 spaces for Malayalam speech to text.

Two options to try it out:

  1. 🤗 spaces

You can try it out in the below link.

  1. GPU-based microservice (coming soon.)

What is behind the name of Vegam and Pallakku?

  • Vegam is a Malayalam word which means speed.
  • Pallakku is a Malayalam word which means a palanquin. It is a vehicle which is carried by people.

As an author of these work, I just wanted to say the meaning of these words doesn’t really have anything to do with the package. Vegam might not be the fastest Malayalam ASR model and Pallakku is named not to refer to the old colonial times. I just named it for the sake of naming it.

Thanks to

  • Santhosh Thottingal for his support and guidance. Santhosh told me about faster-whisper and the utility to convert the model weights to various quantization formats.
  • Organizers of IIIT Summit 23 for giving me an opportunity to present my work.

How was the expereince presenting demos?

Demo presentations are events where you can interact with people during the scheduled time. There was an evaluation panel with one of chief guests of the event Dr. T Ramalingeswara Rao from TCS and one of the professors.

These sessions brought a hacker kind of vibe back to me. I was really happy to receive lot of curious questions and smiles received from folks who came to visit my demo presentation. I set up the demo in such a way that there was a wireless microphone and I would play a few audio files. They used to rate the output of the model.

Some of my professors who taught me in previous semesters like Dr. Keshab Nath visited to see my demo and gave me suggestions. It was a nice vibe to interact with a bunch of BTech students in their second semester from Andhra Pradhesh. Even though they were not able to understand Malayalam, they just hung around and asked me questions :D.