Inference Lab

Applied AI Engineering & Research Lab

Inference Lab is an applied AI research and engineering organization. We develop production-grade AI systems, construct high-quality datasets for underrepresented languages, and publish reproducible research across low-resource NLP, speech intelligence, and AI deployment. Our work is end-to-end: from raw data collection and annotation through model training, evaluation, and deployment as usable software.

---

Research

Low-Resource NLP

Data-Centric Roman Urdu NLP: High-Quality Dataset Curation, Privacy-Preserving Embeddings, and State-of-the-Art Model Benchmarking

A comprehensive data-centric study addressing the critical gaps in Roman Urdu NLP infrastructure. Covers rigorous dataset curation methodology, privacy-preserving embedding strategies, and systematic benchmarking of state-of-the-art models on Roman Urdu classification tasks. Establishes reproducible baselines for future work in this domain.

→ Preprint

---

RUEmoCorp: A Large-Scale Roman Urdu Emotion Corpus with Cross-Institute Annotation Validation and State-of-the-Art Emotion Classification

Construction and release of the largest Roman Urdu emotion recognition corpus to date. Introduces a cross-institute annotation validation framework with structured annotator roles, multi-round calibration, and Inter-Annotator Agreement (IAA) measurement. Accompanies the current state-of-the-art emotion classifier for Roman Urdu.

→ Under Progress

---

Speech AI

Modeling Vocal Fatigue as Embedding-Space Deviation Using Contrastively Trained ECAPA-TDNNs

A novel approach to vocal fatigue detection that frames the problem as deviation measurement in speaker embedding space rather than direct classification. A contrastively trained ECAPA-TDNN encoder is used to capture speaker-specific vocal baselines; fatigue is quantified as geometric distance from the healthy reference embedding. Introduces the ECAPA-TDNN-VHE architecture, achieving 2.5× performance improvement over the standard ECAPA-TDNN baseline.

→ Preprint

---

Continuous Vocal Load Monitoring in Professional Voice Users: Development and Occupational Validation of an Automated Assessment System

A complete occupational health monitoring system for professional voice users — teachers, call center operators, broadcasters, and clinical staff. Addresses the gap between laboratory vocal fatigue research and deployable real-world monitoring tools. Validated against occupational use conditions with a focus on practical deployment in professional environments.

→ Under Review — Journal of Voice

---

Datasets

RUEmoCorp — Largest curated Roman Urdu Emotion Corpus

Multi-class emotion recognition corpus for Roman Urdu, constructed with structured annotation pipelines, cross-institute validation, and rigorous quality control. Supports research in low-resource affective computing and multilingual NLP.

HuggingFace · Harvard Dataverse

---

Roman Urdu Sentiment Corpus — Largest curated Roman Urdu Sentiment Corpus

Large-scale sentiment corpus for Roman Urdu, released with full documentation of collection methodology, annotation schema, and inter-annotator agreement statistics. Serves as the benchmark dataset for Roman Urdu sentiment classification.

HuggingFace · Harvard Dataverse

---

Models

ECAPA-TDNN-VHE — Vocal Health Encoder

Custom ECAPA-TDNN architecture trained contrastively for vocal health assessment. Encodes speaker vocal characteristics into a health-sensitive embedding space. Achieves 2.5× performance improvement over the standard ECAPA-TDNN baseline on vocal fatigue detection benchmarks.

HuggingFace

---

Roman Urdu Emotion Classifier — Current State of the Art

XLM-RoBERTa fine-tuned on RUEmoCorp for multi-class Roman Urdu emotion recognition. Macro F1: 0.9896. The highest-performing publicly available model for this task.

HuggingFace

---

Roman Urdu Sentiment Classifier — Current State of the Art

XLM-RoBERTa fine-tuned on the Roman Urdu Sentiment Corpus. The highest-performing publicly available model for Roman Urdu sentiment classification.

HuggingFace

---

Software

VoiceMonitor Python library for continuous vocal load monitoring. Designed for integration into occupational health workflows, real-time audio pipelines, and professional voice user monitoring systems.

Auralis VFS Vocal fatigue scoring library. Provides a programmable interface for fatigue quantification using the ECAPA-TDNN-VHE encoder. Designed for clinical and occupational deployment scenarios.

VocalID Voice biometrics library for speaker verification and identification. Built for security-sensitive applications requiring speaker authentication from raw audio.

faker-pk Localized synthetic data generation library for Pakistan. Generates realistic dummy data — names, addresses, CNICs, phone numbers, and institutional identifiers — for database seeding, system testing, and privacy-safe development workflows.

---

Standards

Every release from Inference Lab adheres to the following:

• Reproducible training and evaluation pipelines with public code
• Rigorous evaluation reporting — macro F1, per-class metrics, confidence intervals where applicable
• Documented data collection, annotation methodology, and IAA statistics
• Deployable inference code alongside model weights
• Honest documentation of limitations and failure cases
• Archival publication on Harvard Dataverse with permanent DOIs for all datasets

---

Contact

Muhammad Khubaib Ahmad — Founder, Lead Researcher & Engineer

Gmail: inferencelab.ai@gmail.com

GitHub

Multan, Punjab, Pakistan