wave2vec2

Multimodel deepfake detection

Tone.me helps users improve their pronunciation in Mandarin

In this code, we have used common and well-known datasets such as the Toronto dataset available on Kaggle to create a sentiment analysis model from human voice. This model is designed based on the Bert model and is called Hubert.

Adaptive Speech Monitoring for Instruction-Critical Environments

Multimodal Model which take text audio and video to predict the turn taking. That is, to predict whether the speaker in a discussion will change.

Generates section wise topics and transcription for lecture videos and helps to control the lecture video playback based on generated topic-wise timestamps.

Monitor driving instructor speech with real-time keyword spotting and emotion classification to measure instruction quality.

Fine-tuned Wav2Vec2 for Hindi ASR with NLLB-200 translation to English — modular training/eval pipeline, ~30% WER

PyTorch implementation of MixCap: A Multimodal Video Captioning model fusing BLIP-2 (Visual) & Wav2Vec2 (Audio). Features a novel Dual-Target MixUp strategy for low-resource training.

Low-resource multimodal hate speech detection leveraging acoustic and textual representations for robust moderation in Telugu.