Senior ML Engineer – Speech & Voice (NXJ-212)

The platform delivers real-time AI video processing and automated content generation for professional sports leagues globally. It operates in high-noise live broadcast environments requiring tight latency budgets, high throughput, and robust handling of overlapping speech and crowd noise.

• 3–5+ years of experience shipping ML systems into production, with 2+ years dedicated to speech or audio pipelines.
• Deep Python and PyTorch expertise, including hands-on GPU profiling, memory optimization, and inference acceleration.
• Hands-on experience with at least two key speech domains: diarization (pyannote, NeMo/Sortformer, VBx), speaker embeddings (ECAPA-TDNN), or ASR and forced alignment (Whisper, Parakeet, wav2vec-family).
• Strict evaluation rigor: experience building test sets, computing DER/WER/EER correctly, and running statistical hypothesis tests.
• Ability to critically analyze academic papers or model cards and reproduce published claims on internal datasets.
• Hands-on experience deploying containerized GPU services in production environments using Docker and queue/API architectures.

Nice to have:

• Hands-on experience with Azure ecosystem tools (Service Bus, Blob Storage) and event-driven microservices.
• Experience with domain adaptation or fine-tuning speech models on noisy, broadcast, or far-field sports audio.
• Experience handling multilingual STT pipelines, particularly with Hebrew, Arabic, or Spanish.
• Experience designing shadow deployment paths, experiment tracking workflows, and annotation team processes.

• Own and scale the core speech pipeline covering alignment, diarization, speaker identification, and enrollment signature-matching workflows.
• Optimize GPU inference performance for throughput, memory footprint, and cost through batching, precision tuning, and compilation.
• Build and maintain labeled sports test benchmarks stratified by speaker count, audio quality, language, and background noise.
• Define and track production evaluation metrics including DER, WER, word-level speaker attribution, speaker-count error, latency, and GPU cost.
• Conduct rigorous model bake-offs and write clear decision memos detailing trade-offs, licensing constraints, and statistical uncertainty.
• Maintain a shadow or A/B deployment framework to gate every production release with empirical evidence.