Senior Staff Engineer, Automatic Speech Recognition (ASR)
www.gnani.ai
Location: Bengaluru Type: Full-time Function: AI Team
About the role
While you're reading this, Gnani is listening to thousands of customers across India — in Hindi, Tamil, Telugu, and dozens more languages, over noisy phone lines, thick accents, and mid-sentence code-switches. That understanding is our ASR stack. It's the difference between a voice agent that mishears and one that gets it right the first time, in the language and accent the caller actually uses.
We're building the most accurate speech recognition for Indian languages, in real time and at enterprise scale — and we want you to lead the research behind it. As Senior Staff Engineer for ASR, you own the stack end-to-end: advancing accuracy and multilingual coverage, engineering it for streaming latency, and hardening it for the messy realities of production audio. This is a senior individual-contributor role: you set technical direction, design experiments, and mentor the engineers around you — without a people-management load.
Core mandate
Research
Advance ASR research for Indian languages — accuracy, multilingual coverage, and robustness across modern architectures
Production & performance
Ship streaming and offline ASR with low latency, high throughput, and reliability at scale
Voice AI ASR & Feature Enhancements
Handle real-world audio and build the real-time components conversations depend on
Technical leadership
Guide engineers, design rigorous experiments, and own the standard for how ASR gets built here
What you'll drive
Research & modeling
Own the ASR modeling roadmap — architecture selection, accuracy, and multilingual / code-switched coverage across Indian languages
Work across modern ASR architectures (Conformer, E-Branchformer, RNN-T, CTC, Whisper / encoder-decoder, and Speech Language Models) and translate findings into shippable systems
Push on the hard multilingual problems: code-switching, regional accents, dialects, pronunciation variation, and low-resource languages
Train and fine-tune large-scale multilingual models — data collection, augmentation, domain adaptation, contextual biasing, and custom vocabulary
Own the text side of recognition: tokenization (BPE, SentencePiece, phonemes), language modeling, confidence scoring, punctuation, capitalization, and text normalization
Participate in benchmarks and publish the work in international conferences
Production & performance engineering
Design ASR systems for production — streaming (low partial / first-token latency) and offline (high-throughput batch), with favourable RTF
Optimize production inference using ONNX, TensorRT, NVIDIA Triton, quantization, batching, and GPU-efficient serving architectures
Partner with platform and infra teams on serving, scaling, and reliability — latency SLAs, uptime, error budgets
Build evaluation and regression harnesses so accuracy and latency don't silently regress release to release
Robustness & real-time Voice AI
Build robust ASR for real-world environments — noisy audio, telephony, far-field speech, reverberation, and overlapping speakers
Develop the Voice AI components real-time conversations depend on: VAD, endpointing, turn detection, barge-in, speech segmentation, and conversational state management
Turn recurring field failures (accent, domain, channel) into systematic model and pipeline improvements
Technical leadership & research presence
Technically guide engineers — experiment design, tracking, code and research review, and career-shaping mentorship
Translate ASR research into enterprise-scale production systems and define the architecture that gets there
Set the bar for rigor: reproducible experiments, clean baselines, honest evaluation, and clear write-ups
Stay current with ASR research and bring the relevant frontier into Gnani's roadmap
Publish and represent Gnani's ASR work externally — international conferences, talks, and community engagement
Who you are
EXPERIENCE WE'RE LOOKING FOR
8–10 years in speech / ML research or engineering, with substantial contribution to production-grade ASR systems
A track record of designing, training, and deploying streaming and offline ASR at scale — high accuracy, low latency, high reliability
Deep expertise in multilingual ASR, especially Indian languages — code-switching, regional accents, dialects, pronunciation variation, and low-resource languages
Strong command of modern ASR architectures — Conformer, E-Branchformer, RNN-T, CTC, Whisper / encoder-decoder, and Speech Language Models (SLMs)
Good publications at international conferences (e.g. Interspeech, ICASSP, NeurIPS, ICML, ACL) in speech / audio / ML
A proven record of translating ASR research into enterprise-scale production systems and mentoring high-performing engineering teams
WHAT MAKES A STANDOUT CANDIDATE
Experience building ASR for hostile audio — telephony, far-field, reverberation, and overlapping speakers
Hands-on with real-time Voice AI components — VAD, endpointing, turn detection, and barge-in
Depth in tokenization, language modeling, contextual biasing, and text normalization
Familiarity with the Indian voice landscape and low-resource data strategies
Comfortable operating with high ownership in a fast-moving, post–Series B environment
TECHNICAL FLUENCY — A MUST-HAVE
Hands-on with PyTorch, NVIDIA NeMo, ESPnet, Hugging Face, and Kaldi / k2, plus distributed training and large-scale evaluation
Fluency with ASR evaluation — WER (and CER), latency, RTF — and what drives improvements or regressions on each
Production inference optimization — ONNX, TensorRT, NVIDIA Triton, quantization, batching, and GPU-efficient serving
Judgment on evaluation and benchmark design: good test sets, fair comparison across languages / accents / domains / channels, and not overfitting to a benchmark
Why Gnani, why now
Gnani.ai is one of the few companies in India doing original speech research at depth — our models are trained on real Indian language data at scale, not adapted from English-first architectures. As Senior Staff Engineer for ASR, you'll own one of the most technically demanding surfaces in Indian voice AI, with the freedom to set direction and the impact of seeing your models understand real conversations across banking, healthcare, logistics, and government services.
High visibility, high impact, and a clear path to the most senior technical roles as the company scales.
