About the role
About the Opportunity
A fast-moving AI and speech technology startup operating at the intersection of conversational AI, voice analytics, and multilingual ASR systems, we’re building next-gen speech recognition models that serve global enterprises and telecom partners across India and emerging markets. Our team is mission-driven to reduce WER (Word Error Rate) by 30%+ through continuous data curation, acoustic modeling, and fine-tuning — all while scaling to support 20+ Indian languages.
Role & Responsibilities
- Diagnose, analyze and reduce WER across ASR outputs for 10+ Indian languages using error pattern mapping and linguistic tagging.
- Curate, clean, and annotate speech datasets to improve model accuracy — focusing on domain-specific, noisy, or low-resource environments.
- Run controlled A/B tests on acoustic and language models to measure impact of data changes on WER scores.
- Collaborate with ML Engineers to implement fine-tuning strategies for Whisper, Wav2Vec2, or custom transformer-based ASR models.
- Generate detailed WER dashboards and root-cause reports to guide model iteration cycles and prioritize data gaps.
- Document error taxonomies and speech variability patterns (accents, code-switching, background noise) for model training teams.
Skills & Qualifications
- Must-Have
- Python
- ASR Systems
- WER Analysis
- Speech Annotation Tools
- Whisper
- Wav2Vec2
- Acoustic Modeling
- Language Modeling
- Preferred
- Indic Language Phonetics
- PyTorch
- Speech Recognition Evaluation Metrics
Benefits & Culture Highlights
- Flexible remote work with asynchronous collaboration across time zones.
- High-impact role with direct influence on model performance and product quality.
- Opportunity to work with cutting-edge multilingual ASR models and real-world speech datasets.