Research Scientist

Problem Tracks

• Track 1: Automatic Speech Recognition (ASR): Develop end-to-end architectures (Transformer, Conformer) using foundation models like Whisper. Focus on streaming recognition, robustness in noisy/multi-accent environments (including Indian accents), and evaluation metrics like WER/CER.
• Track 2: Text-to-Speech (TTS): Build flow-matching/diffusion-based systems for real-time synthesis. Research controllable accent adaptation, prosody modeling, and voice cloning quality using MOS and speaker similarity metrics.
• Track 3: Speech Enhancement / Noise Suppression: Develop models for low-SNR and reverberant conditions across single/multi-microphone scenarios. Optimize for edge deployment with latency constraints, evaluated on SI-SDR/PESQ/STOI.
• Track 4: Model Optimization: Execute quantization-aware training, pruning, distillation, and pipeline profiling (ONNX/TensorRT) to minimize latency/memory trade-offs on CPU/GPU/NPU targets.

Candidate Profile & Requirements

We seek Research Scientists with 4+ years of hands-on experience in Audio/Speech/Voice AI, Python, PyTorch/TensorFlow. You must possess strong ML/DL fundamentals (CNNs, RNNs, Transformers) and signal processing knowledge.

• Nice to have: C/C++/CUDA skills for optimization tracks; experience with Kaldi, ESPnet, NeMo;

Evaluation Criteria

We value strong fundamentals, research thinking, experimental rigor, and problem-solving ability over mere framework familiarity. Successful candidates will bring new techniques to the team via SOTA literature updates.