AI Engineer - ASR: PraDigi, a Pratham Initiative

Overview

Lead the development and deployment of speech recognition (ASR) systems for automated reading assessment across multiple languages, working across the full ASR pipeline from data collection to deployment.

What you'll do

• Provide technical inputs to speech data collection design, including sampling across languages, dialects, age groups, reading levels, recording environments, devices, and target items.
• Define annotation guidelines and quality checks for transcripts, reading errors, disfluencies, background speech and other phenomena relevant to reading assessment.
• Work with data and field teams to identify data gaps and annotation inconsistencies through model error analysis, and recommend targeted additional data collection or re-annotation.
• Help design annotation schemas and inter-annotator quality processes so labels are consistent enough for model training, alignment, and evaluation.
• Train, fine-tune, and evaluate ASR models for children's read speech across multiple languages and accents.
• Build pipelines for audio preprocessing, segmentation, normalization, transcription alignment, and model training.
• Adapt existing ASR architectures such as Whisper, wav2vec 2.0, Conformer, CTC/RNN-T, or similar models for reading assessment.
• Optimize models for practical deployment environments, including low-bandwidth, store-and-forward, server-based, and potentially offline/mobile inference.
• Develop approaches for detecting substitutions, omissions, insertions, repetitions, partial-word attempts, hesitations and self-corrections.
• Implement forced alignment, phoneme-level alignment, or pronunciation-scoring approaches where required.
• Conduct systematic error analysis by language, reading level, age, accent, gender, recording conditions, and device.
• Collaborate with assessment experts and field teams to translate model outputs into measures such as reading accuracy, WCPM, fluency, and reading level.
• Define and track ASR metrics relevant to reading assessment, including WER/CER, word-level accuracy, error-detection precision/recall, alignment accuracy, and reading-score agreement with human evaluators.
• Build reproducible experimentation and model-evaluation pipelines and maintain clear technical documentation.

What you'll need

• A degree in Mathematics, Computer Science, Data Science or a related STEM field.
• 3-5 years of full-time experience, at least 2 of them in an AI/ML role.
• Strong experience in automatic speech recognition, speech processing, or spoken-language technologies.
• Strong Python skills and experience with PyTorch or similar deep-learning frameworks.
• Experience with modern ASR architectures and toolkits such as Hugging Face Transformers, NeMo, Kaldi, or equivalent.
• Understanding of ASR concepts including acoustic modelling, CTC / transducer / encoder-decoder architectures, decoding and beam search, language models, forced alignment, phoneme representations, and WER/CER and ASR error analysis.
• Experience preparing and working with large-scale audio and transcription datasets.
• Ability to diagnose model errors systematically rather than relying only on aggregate transcription accuracy.

Nice to have

• Ownership of the ASR-based reading assessment pipeline across multiple languages (listed under desired experience).

Details

• Location: Remote, anywhere in India.
• Employment type: Full-time, immediate joining.
• Periodic travel to field locations is expected.
• Employer: Pratham International, hiring through its wholly-owned subsidiary in India (PULS Pvt. Ltd.).