Vaibhav Srivastav’s Post

introducing outeTTS-0.1-350M: text-to-speech without the baggage 🎙️. this sleek model uses pure language modeling, ditching the usual heavy architectures. imagine creating custom voices with just a few seconds of audio - that's zero-shot voice cloning for you! 🤯 with just 350 million parameters, it rivals much larger models. is this the future of tts tech? what do you think about the move towards smaller, smarter models? let's discuss. 👇

Leverage the power of Canary, a multilingual speech-to-text model that ranks at the top of the HuggingFace Open ASR Leaderboard with an average word error rate of 6.67% outperforming all other open-source models by a wide margin. Explore the model and learn how to use it. https://nvda.ws/49ET2rb #SpeechRecognition #Transcription

💎 Overview of Vision-Language Models (#VLM) ✅ Vision-Language Models (#VLMs) integrate both visual (image) and textual (language) information processing. They are designed to understand and generate content that involves both #images and #text, enabling them to perform tasks like #imagecaptioning, #visualquestionanswering, and #texttoimage generation. ✅ This primer offers an overview of their architecture and how they differ from Large Language Models (LLMs). 👉 https://lnkd.in/dnv6425Y ♻️𝒌𝒊𝒏𝒅𝒍𝒚 𝒓𝒆𝒑𝒐𝒔𝒕♻️

Explore the capabilities of Enhanced Natural Language Prompting in Hauska Design Pro. Transform text descriptions into detailed visual designs effortlessly. Simplify your workflow and bring your ideas to life faster than ever. https://hubs.ly/Q02zkkqV0

L2リスニングのコーディングスキーム トップダウン、ボトムアップ、情意的、マルチモーダル、環境的 🔓 Siegel, J., Kuteeva, M., & Siegel, A. (2025). Making sense of non-comprehension issues while listening: A data-based coding scheme. Research Methods in Applied Linguistics, 4(1), Article 100181. https://lnkd.in/gKuXmmhX

Architecture of language: Forced perspective https://lnkd.in/e9yjyFRk

In this episode, we discuss TextSquare: Scaling up Text-Centric Visual Instruction Tuning by @Jingqun Tang, Chunhui Lin, Zhen Zhao, Shu Wei, Binghong Wu, Qi Liu, Hao Feng, Yang Li, Siqi Wang, Lei Liao, Wei Shi, Yuliang Liu, Hao Liu, Yuan Xie, Xiang Bai, @Can Huang. The paper describes advancements in text-centric visual question answering using a novel dataset called Square-10M, developed to improve Multimodal Large Language Models (MLLMs) through instruction tuning. The dataset, generated with closed-source MLLMs, employs a method named Square that covers Self-Questioning, Answering, Reasoning, and Evaluation for data construction. Experiments on the dataset indicated significant performance enhancements over existing models, highlighting the importance of the quantity of reasoning data in VQA for enhancing accuracy and reducing errors in model responses.

Check out our new paper on "Joint Semantic Knowledge Distillation and Masked Acoustic Modeling for Full-band Speech Restoration with Improved Intelligibility"! We introduce a generative model addressing denoising, dereverberation, bandwidth extension, and declipping, producing high-quality full-band (44.1 kHz) speech. Our approach involves masked audio token modeling with a language model. Notably, we propose self-supervised semantic knowledge distillation, which significantly enhances the speech intelligibility without the need for transcribed data nor increasing model size and inference time. Find the paper and samples at the provided link!

GPT-4o Key takeaways 🚀 In the research paper, the human response time across languages-208ms Recognizing and identifying various speakers in an audio recording Targeted alteration of input photos in a Photoshop-like manner Much superior to GPT-4 Turbo in languages other than English Significantly better text rendering on created images and fonts Capacity to communicate human feelings and vocal abilities Benchmarks for MMLU/HumanEval have slightly improved 50% less expensive and twice as quick as GPT-4 Turbo Capability of audio recording human emotions 3D modeling and image creation Summarization of the lecture Read the Analysis here: https://lnkd.in/dGxz-Mhf