Indo-Aryan Indic-Conformer is a multilingual speech model for North-Indian languages. This model is based on Conformer large architecture, with 115M parameters.
2upvotes
106downloads
3,917views
64.91 KB
0Shares
updated 06 Nov 2026
Upvote
2
LICENSE MIT
SECTOR Sector Agnostic
About Model
Bhashini - The Indo-Aryan Indic-Conformer is a multilingual automatic speech recognition (ASR) model designed specifically for North-Indian languages. It is based on the Conformer large architecture, which is known for its efficiency and accuracy in processing speech signals. The model contains 115 million parameters, enabling it to effectively transcribe spoken language into text with high precision.
This ASR model has been trained on the Shrutlip dataset, a rich dataset designed to enhance automatic speech recognition capabilities in Indian languages. The model primarily supports the Odia language and has been developed by AI4Bharat, a leading research initiative focused on advancing AI-driven solutions for Indian languages.
With a batch processing setup, this model is optimized for large-scale speech-to-text tasks across general domains. It is a valuable resource for applications in speech transcription, voice-enabled interfaces, digital accessibility, and natural language processing (NLP) research. Given the increasing demand for multilingual ASR systems, this model serves as a foundational tool for improving speech technology in India’s diverse linguistic landscape.
The Indo-Aryan Indic-Conformer is open-source, and its implementation is available on GitHub, making it accessible for researchers, developers, and AI practitioners working in the domain of Indian language speech processing.
For more details about the use of model, refer to github: https://github.com/AI4Bharat/IndicTrans2/tree/main
Dhīr 5B MoE Q4_K_M GGUF (3.4 GB) for offline llama.cpp in English + 10 Indian languages. 5.09B total/1.79B active parameters; 91% token agreement (1.5M tokens), ≤1.8% perplexity vs BF16.
Phi-3-vision-128k-instruct is a state-of-the-art multimodal model by Microsoft, designed to process both text and visual inputs with a context length of up to 128,000 tokens.
A 14B parameter AI model optimized for 4K and 128K token context lengths, designed for advanced reasoning, math, and coding applications, supporting both research and commercial use.
An ONNX-optimized version of Phi-3-Medium-128K-Instruct, designed for high-speed, efficient inference on NVIDIA GPUs, supporting FP16 and INT4 quantization for enhanced performance.
An ONNX-optimized version of Phi-3-Small-8K-Instruct, designed for efficient inference on NVIDIA GPUs, supporting FP16 and INT4 quantization for fast, scalable AI applications.
A GGUF-optimized version of Phi-3-Mini-4K-Instruct, designed for efficient low-memory AI inference, supporting quantized and full-precision formats for balanced performance and quality.
A 14B parameter AI model optimized for 128K token context length, designed for complex reasoning, mathematical problem-solving, and efficient AI applications in commercial and research environments.
An ONNX-optimized version of Phi-3-Medium-128K-Instruct, designed for efficient long-context inference on CPUs, supporting structured reasoning, text generation, and code processing.
An ONNX-optimized version of Phi-3-Mini-128K-Instruct, quantized for fast, scalable inference on CPUs, GPUs, and mobile devices, supporting long-context AI applications.
An ONNX-optimized version of Phi-3-Medium-4K-Instruct, designed for efficient AI inference on Windows machines using DirectML, supporting INT4 quantization for high-performance execution on AMD, Intel, and NVIDIA GPUs.
A Text-to-Speech (TTS) model developed for the Mizo language that converts written Mizo text into natural-sounding speech, supporting the development of Mizo-language voice applications and accessibility solutions.
A Text-to-Speech (TTS) model developed for the Mizo language that converts written Mizo text into natural-sounding speech, supporting the development of Mizo-language voice applications and accessibility solutions.
This Machine Translation (MT) model translates text from one language to another using a trained neural network. It can support multilingual communication and language translation applications. Translation quality depends on language coverage, training data, and contextual accuracy. For Mizo, evaluate the model using representative test sentences before deployment.
The TTS was built using 5.18 hours of Mizo speech data. The Mizo text-to-speech (TTS) synthesis system built using the Variational Inference with Adversarial Learning for Text-to-Speech (VITS) framework.
Fine-tuned pre trained model for Mizo-English machine translation, trained on Mizo-English parallel datasets to improve translation performance for the Mizo language.
VITS text-to-speech models for Sourashtra, a low-resource Indo-Aryan language. There are male and female voices, and each voice accepts text in either Tamil script or Sourashtra script.
This dataset is designed for training and evaluating Optical Character Recognition (OCR) models capable of extracting text from document images. It contains 14,000 labeled samples consisting of image filenames and their corresponding ground-truth text annotations, along with metadata such as font type, font size, and document category. The dataset includes diverse document styles, including printed books and scanned documents, enabling models to learn robust text recognition under varying visual
Dhīr 5B MoE Q4_K_M GGUF (3.4 GB) for offline llama.cpp in English + 10 Indian languages. 5.09B total/1.79B active parameters; 91% token agreement (1.5M tokens), ≤1.8% perplexity vs BF16.
Phi-3-vision-128k-instruct is a state-of-the-art multimodal model by Microsoft, designed to process both text and visual inputs with a context length of up to 128,000 tokens.
A 14B parameter AI model optimized for 4K and 128K token context lengths, designed for advanced reasoning, math, and coding applications, supporting both research and commercial use.
An ONNX-optimized version of Phi-3-Medium-128K-Instruct, designed for high-speed, efficient inference on NVIDIA GPUs, supporting FP16 and INT4 quantization for enhanced performance.
An ONNX-optimized version of Phi-3-Small-8K-Instruct, designed for efficient inference on NVIDIA GPUs, supporting FP16 and INT4 quantization for fast, scalable AI applications.
A GGUF-optimized version of Phi-3-Mini-4K-Instruct, designed for efficient low-memory AI inference, supporting quantized and full-precision formats for balanced performance and quality.
A 14B parameter AI model optimized for 128K token context length, designed for complex reasoning, mathematical problem-solving, and efficient AI applications in commercial and research environments.
An ONNX-optimized version of Phi-3-Medium-128K-Instruct, designed for efficient long-context inference on CPUs, supporting structured reasoning, text generation, and code processing.
An ONNX-optimized version of Phi-3-Mini-128K-Instruct, quantized for fast, scalable inference on CPUs, GPUs, and mobile devices, supporting long-context AI applications.
An ONNX-optimized version of Phi-3-Medium-4K-Instruct, designed for efficient AI inference on Windows machines using DirectML, supporting INT4 quantization for high-performance execution on AMD, Intel, and NVIDIA GPUs.