Senior ML Research Scientist, Speech

Company and Speech Research

Role and Work Context

Responsibilities

• Lead a research track and set its roadmap with CTO and product leads, tied to accuracy, naturalness, latency, and cost per minute.
• Design architectures and pre-train models from scratch on large multilingual corpora, including self-supervised pre-training, tokenizer and codec design, and scaling experiments.
• Build streaming ASR with code-switch robustness, contextual biasing, and domain adaptation for BFSI and healthcare.
• Build expressive streaming TTS for Indian languages, including voice cloning, prosody and emotion control, and number, date, and mixed-script pronunciation.
• Build full-duplex speech-to-speech models, including audio tokenizers/codecs, speech-LLM pre-training and alignment, turn-taking, and end-of-utterance prediction.
• Set data strategy for sourcing, licensing, synthetic data, labeling quality, and consent-compliant production-audio use.
• Define real-call evaluation benchmarks and compare against external vendors.
• Make serving trade-offs including distillation, quantization, GPU concurrency, and on-premises footprints.
• Mentor 1–3 engineers; review experiment designs and code.
• Publish or open-source selectively where it strengthens the company's position.

Required Qualifications

• 3–6 years in ML, including at least 2–3 years focused on speech, such as ASR, TTS, speaker/voice, or audio-language models.
• MS or PhD in CS, EE, or a related field, or equivalent depth demonstrated through shipped work.
• Pre-trained at least one speech or audio model from scratch that reached production or achieved strong published results; fine-tuning alone is not enough.
• Deep knowledge of modern speech architectures, including the listed ASR, TTS, neural codec, and speech-LLM approaches.
• Large-scale distributed pre-training on multi-node GPU clusters, 100k+ audio hours, scaling laws, compute budgeting, data loading, and debugging unstable runs.
• Strong experimental design, including ablations, significance, and avoiding test-set leakage.
• Clear written and spoken communication with engineers, product teams, and customers.

Preferred Experience

• Publications at Interspeech, ICASSP, ACL/EMNLP, NeurIPS, or similar venues are a plus.
• Indic or low-resource language and code-switching experience is a plus.
• Streaming or real-time speech systems with strict latency SLAs are a plus.
• RL or preference tuning for speech, LLM post-training, and open-source speech toolkit contributions are nice to have.

Success Measures