---
language:
- en
- as
- bn
- brx
- doi
- gu
- hi
- kn
- ks
- kok
- mai
- ml
- mni
- mr
- ne
- or
- pa
- sa
- sat
- sd
- ta
- te
- ur
- bho
- bhb
library_name: transformers
pipeline_tag: automatic-speech-recognition
base_model:
- bodhan-ai/indic-transcribe-core
tags:
- automatic-speech-recognition
- speech
- audio
- asr
- multilingual
- indic
- code-switching
- code-mixing
- language-identification
- canary
- fastconformer
- quantized
- int8
- bitsandbytes
license: other
---

# Indic-Transcribe-Core 8-bit

An **8-bit quantized version** of [Bodhan AI's Indic-Transcribe-Core](https://huggingface.co/bodhan-ai/indic-transcribe-core), created using **BitsAndBytes** quantization.

The model retains the capabilities of the original multilingual ASR model while significantly reducing GPU memory requirements, making it more suitable for resource-constrained inference and serving multiple users.

## Key Features

* 🎙️ Multilingual Automatic Speech Recognition
* 🇮🇳  Supports Indian languages
* ⚡ 8-bit quantization using **BitsAndBytes**
* 💾 Reduced GPU VRAM usage
* 🔌 Designed for efficient inference and deployment

## VRAM Usage

| Model                            | Approx. GPU VRAM |
| -------------------------------- | ---------------: |
| Original `indic-transcribe-core` |            ~5 GB |
| **This 8-bit model**             |      **~2.5 GB** |
| Reduction                        |         **~50%** |

Actual memory usage may vary depending on GPU, batch size, audio duration, and inference configuration.

## Base Model

This model is derived from:

**[bodhan-ai/indic-transcribe-core](https://huggingface.co/bodhan-ai/indic-transcribe-core)**

Please refer to the original model card for the architecture, training details, supported languages, benchmarks, limitations, and licensing information.

## Quantization

The model was quantized to **8-bit using BitsAndBytes**.

The primary goal of this quantized release is to reduce GPU memory consumption while maintaining the capabilities of the original model.



## Inference

For a complete inference example, see the
[Inference Notebook](./transcribe-8bit-quant.ipynb).


## Intended Use

Suitable for:

* Indian-language speech-to-text
* Voice applications
* Multilingual transcription
* Low-VRAM GPU deployment
* Multi-user ASR serving
* On-premise and edge-oriented inference

## Attribution

This is a quantized derivative of **Bodhan AI's Indic-Transcribe-Core**.
Please retain the original model's attribution and comply with its license and usage requirements.

## Disclaimer

This repository contains an **8-bit quantized version** of the original model.
Performance, accuracy, memory usage, and compatibility may vary depending on the inference environment.