Meet the team that turns your ideas into work.

Discover EightBots
EIGHTBOTS
Contact salesSign up
AI provider

AI4Bharat

AI4Bharat is an IIT Madras research lab that develops datasets, tools and trained models for Indian languages. Its work includes language understanding, translation, speech and transliteration.

Explore models

Models & tools

  • Language-specific FastPitch and HiFi-GAN speech synthesis models, with public checkpoint archives.

    Launch date Downloadable

  • Hindi instruction-tuned OpenHathi adaptation trained on AI4Bharat's IndicInstruct collection.

    Launch date not confirmedRestricted accessDownloadable

  • Research speech recognizer adapting a 1.2-billion-parameter Canary architecture to Dehvali Bhili.

    Launch date not confirmedRestricted accessDownloadable

  • 600-million-parameter IndicConformer adaptation trained on about 200 hours of Bhili speech.

    Launch date not confirmedDownloadable

  • Streaming Bhili speech recognizer derived from Nemotron Conformer with selectable decoding latency.

    Launch date not confirmedRestricted accessDownloadable

  • Sarvam-derived LoRA model for Dehvali-Bhili-to-Marathi translation with an agricultural focus.

    Launch date not confirmedDownloadable

  • Sarvam-derived LoRA model for Marathi-to-Dehvali-Bhili translation with agricultural training data.

    Launch date not confirmedDownloadable
  • Bhili TTS

    AI4Bharat · Bhili Models

    Dehvali Bhili speech model fine-tuned from Indic Parler-TTS on ten hours of studio recordings.

    Launch date not confirmedDownloadable

  • Qwen3.5-2B adaptation for reading English Braille from images, trained on synthetic examples.

    Launch date not confirmedDownloadable
  • Cadence

    AI4Bharat · Cadence

    One-billion-parameter punctuation-restoration model for English and 22 Indian languages.

    Launch date not confirmedRestricted accessDownloadable

  • 270-million-parameter punctuation-restoration model for English and 22 Indian languages.

    Launch date not confirmedDownloadable

  • Fully fine-tuned Llama-3.1-8B-Instruct evaluator for Bengali answers against English references.

    Launch date not confirmedDownloadable

  • LoRA-designated Hercule evaluator for Bengali model answers using English references.

    Launch date not confirmedDownloadable

  • French answer evaluator based on full fine-tuning of Llama-3.1-8B-Instruct.

    Launch date not confirmedDownloadable

  • LoRA-designated Hercule model that evaluates French answers using English references.

    Launch date not confirmedDownloadable

  • Fully fine-tuned Hercule model for evaluating German answers against English references.

    Launch date not confirmedDownloadable

  • LoRA-designated Llama-3.1-8B-Instruct adaptation for evaluating German model answers.

    Launch date not confirmedDownloadable

  • Fully fine-tuned Hercule evaluator for Hindi answers against English references.

    Launch date not confirmedDownloadable

  • Hindi answer-evaluation model from Hercule's LoRA-designated releases.

    Launch date not confirmedDownloadable

  • Fully fine-tuned Llama-3.1-8B-Instruct adaptation for evaluating Telugu answers.

    Launch date not confirmedDownloadable

  • LoRA-designated Hercule evaluator for Telugu model answers using English references.

    Launch date not confirmedDownloadable

  • Urdu answer evaluator based on full fine-tuning of Llama-3.1-8B-Instruct.

    Launch date not confirmedDownloadable

  • LoRA-designated Hercule model for evaluating Urdu answers against English references.

    Launch date not confirmedDownloadable
  • Indic Parler-TTS

    AI4Bharat · Indic Parler-TTS

    Joint AI4Bharat and Hugging Face speech model controlled through written voice descriptions.

    Launch date not confirmedRestricted accessDownloadable

  • Pretrained Indic Parler speech model trained on 8,385 hours of multilingual audio.

    Launch date not confirmedRestricted accessDownloadable
  • IndicBART

    AI4Bharat · IndicBART

    Multilingual generation model that maps Indian-language text into shared Devanagari script.

    Launch date not confirmedDownloadable

  • IndicBART generation variant that retains each Indian language's own script.

    Launch date not confirmedDownloadable

  • Separate-script IndicBART adaptation for summarizing text in seven Indian languages.

    Launch date not confirmedDownloadable

  • IndicBART translation model trained on PMI and PIB data for Indian-language-to-English translation.

    Launch date not confirmedDownloadable
  • IndicBERT

    AI4Bharat · IndicBERT

    Original multilingual ALBERT encoder for English and 11 Indian languages.

    Launch date not confirmedRestricted accessDownloadable

  • IndicBERT v2 encoder using back-translated text for translation-language-model training.

    Launch date not confirmedDownloadable

  • Multilingual IndicBERT v2 encoder trained with the masked-language-model objective.

    Launch date not confirmedDownloadable

  • IndicBERT v2 encoder combining masked-language training with Samanantar translation-language modeling.

    Launch date not confirmedDownloadable

  • IndicBERT v2 encoder that unifies Indian-language text into a shared script.

    Launch date not confirmedDownloadable

  • One-billion-parameter bidirectional IndicBERT v3 encoder derived from Gemma.

    Launch date not confirmedRestricted accessDownloadable

  • 270-million-parameter Gemma-derived bidirectional encoder for Indian-language text tasks.

    Launch date not confirmedRestricted accessDownloadable

  • Four-billion-parameter Gemma-derived IndicBERT v3 encoder for Indian-language text.

    Launch date not confirmedRestricted accessDownloadable

  • 600-million-parameter Conformer speech recognizer covering India's 22 scheduled languages.

    Launch date not confirmedRestricted accessDownloadable

  • Assamese speech recognizer with a 120-million-parameter Conformer encoder and hybrid CTC–RNNT decoder.

    Launch date not confirmedRestricted accessDownloadable

  • Bengali speech-recognition model with a 120-million-parameter encoder and hybrid CTC–RNNT decoding.

    Launch date not confirmedRestricted accessDownloadable

  • Bodo speech recognizer with a 120-million-parameter IndicConformer encoder and CTC or RNNT decoding.

    Launch date not confirmedRestricted accessDownloadable

  • Dogri-specific IndicConformer model with a 120-million-parameter encoder and two decoding options.

    Launch date not confirmedRestricted accessDownloadable

  • Gujarati speech-recognition checkpoint with a 120-million-parameter encoder and hybrid CTC–RNNT decoding.

    Launch date not confirmedRestricted accessDownloadable

  • Hindi-specific IndicConformer recognizer with a 120-million-parameter encoder and CTC or RNNT output.

    Launch date not confirmedRestricted accessDownloadable

  • Kannada speech model with a 120-million-parameter IndicConformer encoder and hybrid decoding.

    Launch date not confirmedRestricted accessDownloadable

  • Kashmiri speech-recognition model with a 120-million-parameter encoder and CTC–RNNT decoder.

    Launch date not confirmedRestricted accessDownloadable

  • Konkani-specific IndicConformer recognizer with a 120-million-parameter encoder.

    Launch date not confirmedRestricted accessDownloadable

  • Maithili speech model using a 120-million-parameter encoder with CTC and RNNT decoding options.

    Launch date not confirmedRestricted accessDownloadable

  • Malayalam speech recognizer built on a 120-million-parameter IndicConformer encoder.

    Launch date not confirmedRestricted accessDownloadable

  • Manipuri-specific speech model with a 120-million-parameter encoder and CTC–RNNT decoding.

    Launch date not confirmedRestricted accessDownloadable

  • Marathi IndicConformer recognizer with a 120-million-parameter encoder and two decoder choices.

    Launch date not confirmedRestricted accessDownloadable

  • Nepali speech-recognition checkpoint with a 120-million-parameter Conformer encoder.

    Launch date not confirmedRestricted accessDownloadable

  • Odia speech recognizer with a 120-million-parameter encoder and hybrid CTC–RNNT decoding.

    Launch date not confirmedRestricted accessDownloadable

  • Punjabi-specific IndicConformer model with a 120-million-parameter encoder and CTC or RNNT decoding.

    Launch date not confirmedRestricted accessDownloadable

  • Sanskrit speech-recognition model with a 120-million-parameter encoder and hybrid decoding.

    Launch date not confirmedRestricted accessDownloadable

  • Santali IndicConformer checkpoint with a 120-million-parameter encoder and CTC–RNNT decoder.

    Launch date not confirmedRestricted accessDownloadable

  • Sindhi speech recognizer with a 120-million-parameter encoder and two decoding options.

    Launch date not confirmedRestricted accessDownloadable

  • Tamil speech-recognition model with a 120-million-parameter Conformer encoder and hybrid CTC–RNNT decoding.

    Launch date not confirmedRestricted accessDownloadable

  • Telugu-specific IndicConformer speech model with a 120-million-parameter encoder.

    Launch date not confirmedRestricted accessDownloadable

  • Urdu speech-recognition checkpoint with a 120-million-parameter encoder and CTC–RNNT decoding.

    Launch date not confirmedRestricted accessDownloadable
  • IndicF5

    AI4Bharat · IndicF5

    Speech-generation model for 11 Indian languages using text, reference audio and its transcript.

    Launch date not confirmedRestricted accessDownloadable

  • IndicBERT-derived component for language identification in Romanized Indian text.

    Launch date not confirmedDownloadable

  • IndicLID fastText component for identifying languages written in their native scripts.

    Launch date not confirmedDownloadable

  • IndicLID fastText component for language identification in Romanized text.

    Launch date not confirmedDownloadable
  • IndicNER

    AI4Bharat · IndicNER

    Multilingual BERT-derived entity recognizer for 11 Indian languages, trained on Samanantar-derived data.

    Launch date not confirmedRestricted accessDownloadable
  • IndicSeamless

    AI4Bharat · IndicSeamless

    SeamlessM4T v2 adaptation for Indian-language speech-to-text translation, trained on BhasaAnuvaad.

    Launch date not confirmedRestricted accessDownloadable

  • Original Transformer-4x translation family trained on Samanantar for English–Indic and Indic–Indic directions.

    Launch date not confirmedAccess not confirmed

  • 1.1-billion-parameter translation model from English into 22 Indian languages.

    Launch date not confirmedRestricted accessDownloadable

  • 200-million-parameter-class distilled model for English-to-Indian-language translation.

    Launch date not confirmedRestricted accessDownloadable

  • Billion-parameter-class IndicTrans2 model for translation from 22 Indian languages into English.

    Launch date not confirmedRestricted accessDownloadable

  • 200-million-parameter-class distilled IndicTrans2 model for Indian-language-to-English translation.

    Launch date not confirmedRestricted accessDownloadable

  • Billion-parameter-class IndicTrans2 model for translation between Indian languages.

    Launch date not confirmedRestricted accessDownloadable

  • 320-million-parameter-class distilled translation model for Indian-language pairs.

    Launch date not confirmedRestricted accessDownloadable

  • Gemma 3 adaptation for sentence and document translation with varying support across Indian languages.

    Launch date not confirmedRestricted accessDownloadable

  • Nine language-specific Fairseq speech recognizers with separate dictionaries, lexicons and decoding models.

    Launch date not confirmedDownloadable

  • Original base-sized IndicWav2Vec speech-pretraining checkpoint for downstream adaptation.

    Launch date not confirmedDownloadable

  • Original large-sized IndicWav2Vec speech-pretraining checkpoint, separate from language-specific recognizers.

    Launch date not confirmedDownloadable

  • Twelve language-specific Whisper adaptations trained on Vistaar speech data.

    Launch date not confirmedDownloadable

  • IndicXlit v1.0 transliteration checkpoint from Indian scripts to Romanized text.

    Launch date not confirmedDownloadable

  • IndicXlit v1.0 transliteration checkpoint from Romanized text to Indian scripts.

    Launch date not confirmedDownloadable

  • English answer evaluator adapted from Llama-3.1-8B-Instruct using Feedback-Collection and Prometheus 2 training.

    Launch date not confirmedDownloadable

  • 120-million-parameter self-supervised multilingual Conformer encoder for speech features.

    Launch date not confirmedDownloadable

  • Unified-script IndicBART model for headline generation in 11 Indian languages.

    Launch date not confirmedDownloadable

  • Separate-script IndicBART model trained on 1.316 million paragraphs for headline generation in 11 Indian languages.

    Launch date not confirmedDownloadable

  • Separate-script IndicBART adaptation for paraphrasing text in 11 Indian languages.

    Launch date not confirmedDownloadable

  • Unified-script IndicBART adaptation for paraphrase generation in 11 Indian languages.

    Launch date not confirmedDownloadable

  • Question-generation model for 11 Indian languages that retains separate scripts.

    Launch date not confirmedDownloadable

  • Unified-script IndicBART model for question generation in 11 Indian languages.

    Launch date not confirmedDownloadable

  • Separate-script IndicBART model for sentence summaries in 11 Indian languages.

    Launch date not confirmedDownloadable

  • Unified-script IndicBART model for sentence summarization in 11 Indian languages.

    Launch date not confirmedDownloadable

  • Separate-script IndicBART adaptation for biography generation in nine Indian languages.

    Launch date not confirmedDownloadable

  • Unified-script IndicBART adaptation for biography generation in nine Indian languages.

    Launch date not confirmedDownloadable

  • Llama 2 research adaptation with instruction training directly on native-script data.

    Launch date not confirmedResearch releaseDownloadable

  • Llama 2 research adaptation with instruction training directly on Romanized data.

    Launch date not confirmedResearch releaseDownloadable

  • RomanSetu research series with five checkpoints of continued pretraining on native-script text.

    Launch date not confirmedResearch releaseDownloadable

  • RomanSetu research series with five checkpoints of continued pretraining on Romanized Indian-language text.

    Launch date not confirmedResearch releaseDownloadable

  • RomanSetu Llama 2 adaptation with native-script continued pretraining followed by native-script instruction tuning.

    Launch date not confirmedResearch releaseDownloadable

  • RomanSetu Llama 2 adaptation with Romanized continued pretraining followed by Romanized instruction tuning.

    Launch date not confirmedResearch releaseDownloadable

  • Sarvam Translate LoRA research adaptation for Bhili-to-English translation.

    Launch date not confirmedResearch releaseDownloadable

  • Sarvam Translate LoRA research adaptation for Bhili-to-Marathi translation.

    Launch date not confirmedResearch releaseDownloadable

  • Sarvam Translate LoRA research adaptation for English-to-Bhili translation.

    Launch date not confirmedResearch releaseDownloadable

  • Sarvam Translate LoRA research adaptation for Marathi-to-Bhili translation.

    Launch date not confirmedResearch releaseDownloadable

  • Multilingual VITS speech model with predefined voices and styles for 13 Indian languages.

    Launch date not confirmedRestricted accessDownloadable