Aditya Kumar Singh

Speech AI

ASR · TTS · Language Identification · Speaker Recognition · Speech Quality Evaluation

Applied ML

Fine-tuning · Experimentation · Benchmarking · Human-in-the-loop Validation · Statistical Evaluation

LLM Systems

LLM APIs · Context-aware prompting · Linguistic edge cases · Embeddings · AI pipelines

Production Engineering

FastAPI · REST APIs · Docker · PostgreSQL · AWS · GCP · Linux

VOICE AI EVALUATION & BENCHMARKING

Designed experiments and statistical evaluation methodologies for benchmarking TTS and ASR systems, formulating hypotheses, conducting significance tests, and analyzing experimental results to identify meaningful differences in model performance.

Language - Gender Identification & Speaker Recognition

Fine-tuned a wav2vec XLS-R model for language and gender identification across 13 Indian regional languages and built speaker/demographic classification using ECAPA voice embeddings and SVM.

Confidential — built @ Josh Talks

Indic ASR Fine-tuning & Evaluation

Fine-tuned Whisper, Wav2Vec2 and NVIDIA NeMo ASR models for Indic datasets, using multi-GPU PyTorch DDP for large-scale experimentation, benchmarking and hyperparameter optimization.

Voice → AI → Form Filling

An AI-powered voice interface for filling service-counter forms, combining speech recognition, backend APIs and an interactive frontend with multilingual support as the product direction.

Private repository

Applied AI Engineer — Josh Talks

• Building production-grade multilingual speech AI systems for Indic ASR and TTS applications, with emphasis on inference optimization, evaluation infrastructure and scalable audio workflows.
• Core contributor to Voice Arena, an automated speech-quality evaluation system spanning MOS, CMOS, MUSHRA, WER/CER, pronunciation verification and statistical validation.
• Engineered context-aware LLM prompts for linguistic edge cases including lexical vs. numeric ordinals and Hindi transliteration while preserving transcript alignment.

AI Research Intern — Josh Talks

• Fine-tuned Whisper, Wav2Vec2 and NVIDIA NeMo ASR models for Indic datasets using PyTorch DDP across multi-GPU environments.
• Built automated experimentation, benchmarking, human-in-the-loop validation and research-analysis workflows.
• Worked with torchaudio and Librosa for audio preprocessing, waveform normalization, spectrogram analysis and multilingual speech validation.

Machine Learning Intern — TechtoGreen Drone & Robotics

• Engineered a Pearson-correlation-driven LSTM-Attention forecasting architecture and data pipeline, reducing RMSE by 12% over baseline models.
• Led the finalist team of the Bhashini Domain Innovation Challenge among 200+ competing teams.