Responsibilities: • Fine-tune and adapt open-source speech models on our proprietary call audio • Build training and evaluation pipelines for multilingual and code-mixed speech • Own model quality against metrics that matter operationally — entity and numeric accuracy, latency to first response — not just aggregate error rates • Design and run data curation at scale: pseudo-labelling, speech enhancement, quality filtering on messy real-world audio • Work with our linguist on text normalisation and pronunciation handling • Evaluate candidate architectures, make the call with evidence, and ship the result to production with the platform team Requirements Must have skills: • 3–4 years in ML, with at least 18 months on speech or audio specifically • Strong Python and PyTorch; comfortable reading a paper and implementing it • Hands-on experience fine-tuning at least one production speech model • Solid grasp of speech fundamentals — mel-spectrograms, acoustic models and vocoders, encoder-decoder vs transducer architectures, evaluation methodology, sampling rates and what they cost you • Understanding of how modern speech systems are actually built: self-supervised encoders, neural audio codecs, LM-based generation, flow matching • Experience with genuinely messy audio, not only clean benchmark datasets Nice to have: • NeMo, ESPnet, SpeechBrain, or Coqui • Telephony-band or contact-centre audio • Multilingual or code-switched speech work • LoRA/PEFT, distributed training • Open-source contributions or publications in speech