About Humyn Labs

At Humyn Labs, we believe the best AI is built on the best human judgment. We operate a global network of 1M+ verified experts who deliver high-quality, multimodal training datasets across domains — backed by reputation verification and multi-layer quality control.

Humyn Labs converts human action — across sound, sight, movement, and touch — into high-quality multi-modal data signals for physical AI. Operating across 20+ countries in India, southeast Asia, Latin America, and the Middle East: the real-world environments where physical AI deploys, not the labs where it is built.

Our data isn't just collected; it's evaluated, defended, and production-ready. Because before AI can be trusted, its training data must be.

Our work sits at the intersection of egocentric video understanding, embodied AI, robotics perception, and voice-driven interaction. We move fast, obsess over data quality, and ship at scale.

Role Overview

We are building structured, high-quality voice datasets for frontier AI companies working on speech-to-text, speech-to-speech, and multimodal AI systems.

We are looking for a Machine Learning Researcher with a focus on voice and speech AI — someone who can rigorously evaluate datasets across evolving speech models, identify performance gaps across Indic and global languages, and publish those findings as structured research for the broader AI community.

This role sits at the intersection of benchmarking, linguistic diversity, and data strategy. If you are deeply curious about how models fail — especially across underrepresented languages and accents — this is built for you.

What You Will Work On

Cross-Model Benchmarking & Evaluation

• Benchmark voice datasets across ASR and speech models (Whisper, Deepgram, Google STT, Azure Speech, and emerging open-source models)
• Measure performance using WER, CER, MOS, robustness, latency, and error pattern analysis
• Design structured experiments to understand how dataset characteristics impact model accuracy
• Compare performance across multilingual, dialect-heavy, emotional, and noisy speech data

Model Gap Analysis — Indic & Global Languages

• Systematically identify where speech models underperform across: Indic languages and dialects (Hindi, Tamil, Telugu, Bengali, Kannada, etc.), code-switching and transliteration, emotional and conversational speech, low-resource language scenarios, and background noise / real-world audio conditions
• Quantify model weaknesses through structured, reproducible analysis
• Map performance gaps to specific dataset requirements — you will help define what data models actually need next

Dataset Quality & Supplier Scoring

• Build a standardized dataset quality scoring rubric with measurable criteria: audio clarity, speaker diversity, annotation accuracy, emotion depth, and accent/dialect coverage
• Tag and rank data suppliers based on objective quality signals

Research Publishing & Community Presence

• Publish benchmarking findings as blog posts and LinkedIn articles accessible to both technical and non-technical audiences
• Contribute to internal evaluation reports tracking performance shifts as new models are released
• Stay current on evolving speech model architectures and share outside-in insights with AI research teams and clients

You Must Have

• 1–3 years of experience in speech AI, audio ML, NLP, or applied AI research
• Hands-on experience with ASR/TTS systems and understanding of model behaviour
• Exposure to running experiments, evaluating models, and designing evaluation frameworks
• Strong Python skills and comfort with ML experimentation workflows
• Genuine interest in linguistic diversity — particularly Indic languages — and how models perform across them
• Strong written communication skills with the ability to turn research into clear, publishable content

Technical Skills

• Python (mandatory)
• PyTorch or TensorFlow
• Whisper, SpeechBrain, Kaldi, or similar toolkits
• Familiarity with WER, CER, MOS, SNR metrics
• Experience with multilingual or low-resource datasets (preferred)

Ideal Mindset

• Curious about model failure modes, not just model capabilities
• Analytical and detail-oriented, with a bias for reproducibility
• Comfortable reading research papers and independently testing new APIs
• Excited to share work publicly — blogs, LinkedIn, open datasets