Function : Data Science & Analytics

As a Senior ML Research Scientist (speech), you will own an entire research direction (STT, TTS, or speech-to-speech) end-to-end. You will design novel neural architectures, lead foundation model pre-training from scratch on massive multilingual audio datasets (100k+ hours), and ship models to production while mentoring engineers on the team.

Responsibilities :

- Lead a Research Track : Define and execute the roadmap for STT, TTS, or S2S alongside the CTO and product leads, tying research outcomes to business metrics (accuracy, naturalness, latency, cost-per-minute).

- Foundation Pre-training at Scale : Design novel architectures and pre-train foundation speech models from scratch on massive multilingual corpora (100k+ audio hours), including self-supervised learning, tokenizer/codec design, and scaling laws.

- STT Innovations : Develop streaming ASR with low time-to-first-token, robust code-switching handling, contextual biasing for enterprise entities, and BFSI/healthcare adaptation.

- TTS Innovations : Build expressive, low-latency streaming TTS for Indian languages with voice cloning, prosody/emotion control, and accurate pronunciation of mixed-script text, numbers, and dates.

- Full-Duplex Speech-to-Speech : Build conversational voice models in-house: audio tokenizers, neural codecs (EnCodec, DAC, Mimi), speech-LLM alignment, turn-taking, and interruption handling.

- Data Strategy : Lead audio sourcing, licensing, synthetic data generation, annotation quality, and consent-compliant usage of real production audio.

- Platform and Serving Trade-Offs : Partner with the platform team on model distillation, quantization, GPU concurrency, and on-prem deployment footprints.

- Mentorship and Publishing : Mentor 1 - 3 engineers, review experiment designs/code, and publish/open-source selectively to advance the company's research standing.

Requirements :

- 3 - 6 years in machine learning, with at least 2 - 3 years focused strictly on speech (ASR, TTS, speaker/voice, or audio-language models).

- Degree : MS or PhD in Computer Science, Electrical Engineering, or a related field (or equivalent depth demonstrated through shipped production models).

- Track Record of Pre-training : Proven experience pre-training at least one speech foundation model from scratch (not just fine-tuning) that reached production or published benchmarks.

- Deep Architectural Knowledge : Conformer/Zipformer, RNN-T/TDT, encoder-decoder ASR, flow-matching and diffusion TTS, neural codecs (EnCodec, DAC, Mimi), and speech-LLMs.

- Large-Scale Distributed Pre-training : Hands-on experience with multi-node GPU clusters using FSDP, DeepSpeed, or Megatron on 100k+ audio hours; experienced in compute budgeting, scaling laws, and debugging loss spikes/unstable runs.

- Scientific Rigor : Strong experimental design (ablations, statistical significance, leakage prevention) and benchmark design against external vendors.

- Leadership and Communication : Clear technical communication across engineering, product, and leadership, with prior experience mentoring junior engineers.

Nice to Have :

- First-author or co-author publications at Interspeech, ICASSP, ACL/EMNLP, NeurIPS, ASRU, or similar.

- Deep experience with Indic languages, low-resource acoustics, or Hinglish code-switching.

- Production experience building real-time speech systems under strict latency SLAs (< 100ms).

- Experience with RL or preference tuning for speech (naturalness rewards) or LLM post-training.

- Contributions to open-source speech toolkits (NeMo, ESPnet, K2/icefall).