Senior Research Scientist, Lilt

• Strong proficiency in Python and experience with popular OSS machine learning frameworks (e.g., PyTorch, TensorFlow)
• Hands-on experience with ASR/OCR open-source toolkits (e.g., Kaldi, Vosk, Whisper, Tesseract)
• Deep understanding of speech signal processing and noise-robust ASR techniques
• Familiarity with speaker diarization, source separation, and audio preprocessing methods
• Experience in deploying production-grade ML systems at scale
• Strong problem-solving skills and ability to work with ambiguous, noisy datasets

Desirable

• We’re looking for a Senior Research Scientist to join our Research & Engineering team, with a focus on Optical Character Recognition (OCR), Automatic Speech Recognition (ASR), and open-source systems (OSS)
• We are seeking a highly skilled Senior Research Engineer to design, develop, and productionize cutting-edge OCR (Optical Character Recognition) and ASR (Automatic Speech Recognition) systems leveraging open-source (OSS) models
• In this role, you will design, develop, and deploy production-grade prototypes that leverage open-source AI models to solve complex real-world challenges — including poor-quality audio and noisy, uncontrolled recording environments
• You’ll collaborate closely with our applied research scientists, product engineers, and linguists to push the boundaries of what’s possible in document and speech understanding
• You will work on challenging real-world problems, including building robust pipelines for poor-quality audio, multi-speaker scenarios, and complex noise environments
• Your work will directly advance the capabilities of speech and text recognition systems in demanding, uncontrolled conditions
• Prototype Development: Design and implement production-grade prototypes for OCR and ASR systems based on open-source models
• Audio Robustness Engineering: Build ASR systems resilient to severe audio challenges such as low/high volume, distortions, overlapping speech, unknown speaker counts, and off-axis microphone placement
• Model Post-Training: Fine-tune and post-train ASR and OCR models for domain-specific accuracy improvements
• Speaker Identification: Develop robust OSS-based solutions for accurately identifying speakers in multi-speaker environments
• Noise Identification: Create systems capable of detecting both speech and environmental noises (e.g., rattling keys, doors closing) within recordings
• Speaker Separation: Implement pipelines to isolate and export individual speaker audio into separate files
• Speech Segmentation: Engineer accurate segmentation of speech into logical and temporal units for downstream processing