About the role
Responsibilities:
- Train speech-to-text (STT) and text-to-speech (TTS) models from scratch.
- Adapt models for Indian languages, Hinglish, and banking/insurance vocabulary.
- Build parts of speech-to-speech systems (tokenizers, codecs, streaming decoders).
- Build data pipelines: cleaning, labeling, and augmentation.
- Evaluate models (WER, entity accuracy, latency) and optimize them for deployment.
- Analyze production errors and turn them into test sets.
- Senior level (3+ years): Own a research area (STT, TTS, or speech-to-speech), pre-train large models on 100k+ hours of audio, define benchmarks, and mentor 1-3 engineers.
Requirements:
- 1-6 years in ML, with at least one real speech or audio project.
- B. Tech/M. Tech/MS in CS, EE, or a related field.
- Strong Python and PyTorch; able to build a model and training loop yourself.
- Speech basics: spectrograms, mel features, CTC, seq2seq.
- Experience with one of NeMo, ESPnet, Hugging Face, fairseq, or K2/icefall.
- Trained at least one model from scratch (multi-GPU, mixed precision, W& B/MLflow).
- Careful evaluation habits.
For 3+ years, also:
- 2-3+ years in speech.
- Pre-trained a speech or audio model from scratch that reached production.
- Distributed training experience (FSDP/DeepSpeed, multi-node).
Nice to have:
- Hindi or another Indian language.
- Publications or open-source work.
- Real-time or telephony audio experience.