Senior Data Scientist - ASR

Location: Bengaluru, India

Department: Engineering.

Experience: 3 - 6 Years

• Train the next Parrotlet-A models end-to-end. Multimodal LLM-based ASR is the core: audio-encoder + projector + LLM-decoder models trained through modality alignment, instruction tuning and RL post-training (GRPO, DPO) on clinical-entity rewards. Alongside, CTC / RNN-T / TDT models (Parakeet, Conformer families) for low-latency streaming, all on multi-thousand-hour medical and Indic corpora.
• Win on drug names and clinical entities. Extend the contextual-biasing stack (G2P phonetic matching, WFST biasing, shallow fusion with medical LMs) so that medicines, dosages and investigations are transcribed right the first time.
• Ship to production. Serve models on GPUs: quantisation, distillation, batching, latency/throughput tuning, and A/B evaluation inside EkaScribe.
• Own the data and eval engine. Build and maintain the data pipelines, test sets and evaluation framework that tell us whether a model is ready for doctors.
• Publish and open-source. Release model cards, benchmarks and papers under Apache 2.0.
• Raise the bar for the team. Review experiments, mentor engineers, and turn research findings into reusable training recipes.

• 2-6 years building ASR or speech models that reached real users, with at least one model you trained from data to deployment.
• Deep, hands-on fluency in one modern speech toolkit and the architectures behind it: CTC, transducers, attention encoder-decoders, SSL encoders, and speech-LLMs. You have aligned an audio encoder to an LLM, or fine-tuned one (Qwen-Audio, Whisper-encoder + LLM, Canary, Gemma-audio style), and know where they beat and where they lose to transducers.
• Strong PyTorch and multi-GPU training skills: mixed precision, distributed data loading, checkpoint surgery, and the instinct to find why a run is slow or diverging.
• Comfort with the full evaluation story: WER/CER and why they lie, entity-level metrics, calibration, and designing test sets that predict production behaviour.
• Working knowledge of audio fundamentals (resampling, codecs, VAD, noise and reverberation) and of how speech data gets messy in the field.
• Clear written communication: you can turn an experiment into a one-page finding others can act on.

• Code-mixed or low-resource Indic language ASR, including familiarity with Devanagari / Dravidian script transliteration.
• Contextual biasing, hotword boosting or LM fusion in production.
• RL or preference-based post-training for sequence models; hallucination control and timestamping in LLM-based ASR.
• Model compression and serving: TensorRT, ONNX, CTranslate2, Triton, vLLM-style batching.
• Publications or open-source contributions in speech; prior exposure to healthcare, call-centre or legal dictation domains.

• Scale that matters. Our models already process consultations for doctors across India; a point of WER here changes what goes into a patient’s record.
• Compute and data you will not find elsewhere. Multi-thousand-hour in-house clinical speech corpora, a huge text corpus for LM fusion, and dedicated GPU capacity under the IndiaAI Mission.
• Open by default. The Parrotlet family ships under Apache 2.0 / MIT; your work carries your name, in model cards and in papers.
• A small, senior team. You sit with the people building our vision, embeddings and clinical-LLM models, reporting directly to the CDSO, with no layers between an idea and a training run.