Voice AI researcher and Senior Research Scientist working on full-duplex spoken dialogue, turn-taking, and speech LLMs.https://shangeth.com/https://shangeth.com/posts/llmranker/https://shangeth.com/posts/llmranker/LLM-based ranking and reasoning algorithms for search and recommendation: why asking an LLM to rank a shortlist outperforms embedding distance on compositional queries, a research-grounded tour of pointwise, pairwise, listwise, setwise, and tournament-style ranking, and a package that implements all of them over any provider.Tue, 04 Aug 2026 00:00:00 GMThttps://shangeth.com/posts/dualturn-production/https://shangeth.com/posts/dualturn-production/Taking DualTurn from an Interspeech 2026 paper into a live voice agent: a ~1.5M-parameter model trained with zero manual labels, fastest end-to-end in real-call testing against cloud and open-source turn detectors, the only one in the industry that tells a backchannel from a real interruption, and set up to keep improving for free from every call it handles.Fri, 24 Jul 2026 00:00:00 GMThttps://shangeth.com/posts/turn-taking-human-vs-ai/https://shangeth.com/posts/turn-taking-human-vs-ai/The checklist for making a voice agent "sound human": less interrupting, more backchanneling, an even exchange, mostly targets things that aren't broken. Measured against the right human (task dialogue, not casual chat), almost everything holds up except response speed, and fixing that carelessly breaks the one thing the agent already gets right.Thu, 25 Jun 2026 00:00:00 GMThttps://shangeth.com/posts/dualturn/https://shangeth.com/posts/dualturn/Speech-to-speech models know when to speak but can't reason. Cascaded LLM pipelines can reason but only react to silence. DualTurn pretrains on dual-channel human conversation to bring S2S-level turn-taking into a standard ASR-LLM-TTS stack.Fri, 03 Apr 2026 00:00:00 GMThttps://shangeth.com/publications/dualturn/https://shangeth.com/publications/dualturn/Dual-channel generative pretraining for learning natural turn-taking in spoken dialogue without labeled data. A 0.5B model that outperforms models 6x its size on turn prediction.Mon, 09 Mar 2026 00:00:00 GMThttps://shangeth.com/posts/speechllm/https://shangeth.com/posts/speechllm/A small multimodal LLM that reads paralinguistic signal — emotion, prosody, speaker traits — directly from speech audio instead of through an ASR transcript, built alongside the release of SpeechLLM at Skit.ai.Wed, 26 Jun 2024 00:00:00 GMThttps://shangeth.com/posts/speech-llms-conversations/https://shangeth.com/posts/speech-llms-conversations/A multimodal speech LLM that processes audio directly to enhance conversational AI while reducing overhead compared to traditional ASR-LLM-TTS pipelines.Thu, 09 May 2024 00:00:00 GMThttps://shangeth.com/publications/slu-prosodic/https://shangeth.com/publications/slu-prosodic/Two techniques for incorporating prosody into end-to-end SLU: prosody-attention and prosody-distillation. Up to 8% intent classification accuracy improvement on SLURP.Sun, 20 Aug 2023 00:00:00 GMThttps://shangeth.com/publications/icassp2023/https://shangeth.com/publications/icassp2023/Map-Mix: a data augmentation approach using model training dynamics to guide latent mixup sampling, giving ~2% weighted F1 improvement on low-resource dialect classification.Sun, 04 Jun 2023 00:00:00 GMThttps://shangeth.com/publications/skit-s2i/https://shangeth.com/publications/skit-s2i/The first public Indian-accented SLU dataset in the banking domain. SSL speech representations beat ASR-based approaches for intent classification.Mon, 26 Dec 2022 00:00:00 GMThttps://shangeth.com/posts/feature-disentanglement/https://shangeth.com/posts/feature-disentanglement/How deep learning models can isolate independent factors of variation in data through VAEs and Beta-TCVAE, enabling controlled synthesis and better downstream representations.Tue, 22 Feb 2022 00:00:00 GMThttps://shangeth.com/publications/speaker-representation/https://shangeth.com/publications/speaker-representation/A semi-supervised framework for speaker profiling that leverages external unlabelled corpora via supervised, unsupervised, and consistency training, achieving RMSE of 6.8 years on age estimation.Sun, 24 Oct 2021 00:00:00 GMThttps://shangeth.com/posts/code-mixing-seminar/https://shangeth.com/posts/code-mixing-seminar/Notes from a seminar covering six papers on code-mixing across NLP, speech synthesis, and speech recognition — including multilingual synthesis and code-mixed ASR.Tue, 24 Aug 2021 00:00:00 GMThttps://shangeth.com/posts/kl-divergence/https://shangeth.com/posts/kl-divergence/A walkthrough of information entropy, KL divergence, mutual information, and cross entropy — with PyTorch implementations.Tue, 01 Sep 2020 00:00:00 GMThttps://shangeth.com/posts/off-policy-monte-carlo/https://shangeth.com/posts/off-policy-monte-carlo/How importance sampling lets us estimate value functions under a target policy using episodes collected by a different behavior policy.Sat, 01 Aug 2020 00:00:00 GMThttps://shangeth.com/publications/pmlr/https://shangeth.com/publications/pmlr/Design and results of the AutoDL challenge series 2019 (AutoCV, AutoCV2, AutoNLP, AutoSpeech, AutoDL), showing winning solutions generalize to unseen datasets.Mon, 01 Jun 2020 00:00:00 GMThttps://shangeth.com/publications/ads/https://shangeth.com/publications/ads/A novel generic mathematical formulation of AutoML unifying HPO and meta-learning, showing meta-learning addresses AutoML more fundamentally than hyperparameter optimization.Mon, 16 Sep 2019 00:00:00 GMThttps://shangeth.com/posts/gan-5/https://shangeth.com/posts/gan-5/How Conditional GANs extend vanilla GANs to generate class-specific samples, and how Pix2Pix does image-to-image translation.Sat, 25 May 2019 00:00:00 GMThttps://shangeth.com/posts/gan-4/https://shangeth.com/posts/gan-4/Implementing DCGAN with CNNs, batch normalization, and transposed convolutions to generate Street View House Numbers.Fri, 24 May 2019 00:00:00 GMThttps://shangeth.com/posts/gan-2/https://shangeth.com/posts/gan-2/How the adversarial game between Generator and Discriminator works, and why equilibrium matters.Tue, 21 May 2019 00:00:00 GMThttps://shangeth.com/posts/gan-3/https://shangeth.com/posts/gan-3/Step-by-step PyTorch implementation of a vanilla GAN trained on MNIST — data loading, Discriminator, Generator, training loop.Tue, 21 May 2019 00:00:00 GMThttps://shangeth.com/posts/gan-1/https://shangeth.com/posts/gan-1/What GANs are, how they work, and some remarkable recent research — StackGAN, iGAN, Pix2Pix.Mon, 20 May 2019 00:00:00 GMThttps://shangeth.com/posts/unsupervised-learning/https://shangeth.com/posts/unsupervised-learning/An intro to unsupervised learning — clustering, feature learning, and dimensionality reduction, and why it matters.Mon, 20 May 2019 00:00:00 GMThttps://shangeth.com/publications/icacds/https://shangeth.com/publications/icacds/A data-driven deep learning approach combining CNN feature extraction with Neural Arithmetic Logic Units (NALU) for stock price prediction using historical price data.Fri, 12 Apr 2019 00:00:00 GMT