Role Summary

Centific AI Research seeks a Research Intern to design and evaluate speech‑first models, with a focus on Spoken Language Models (SLMs) that reason over audio and interact conversationally. You’ll move ideas from prototype to practical demos, working with scientists and engineers to deliver measurable impact.

Scope of Work

• End‑to‑end speech dialogue systems (speech‑in/speech‑out) and speech‑aware LLMs.
• Alignment between speech encoders and text backbones via lightweight adapters.
• Efficient speech tokenization and temporal compression suitable for long‑form audio.
• Reliable evaluation across recognition, understanding, and generation tasks—including robustness and safety.
• Latency‑aware inference for streaming and real‑time user experiences.

Example Projects

• Prototype a conversational SLM using an SSL speech encoder and a compact adapter on an existing LLM; compare against strong baselines.
• Create a data recipe that blends conversational speech with instruction‑following corpora; run targeted ablations and report findings.
• Build an evaluation harness that covers ASR/ST/SLU and speech QA, including streaming metrics (latency, stability, endpointing).
• Ship a minimal demo with streaming inference and logging; document setup, metrics, and reliability checks.
• Author a crisp internal write‑up: goals, design choices, results, and next steps for productionization.

Minimum Qualifications

• Masters or PhD candidate in CS/EE (or related) with research in speech, audio ML, or multimodal LMs.
• Fluency in Python and PyTorch, with hands‑on GPU training; familiarity with torchaudio or librosa.
• Working knowledge of modern sequence models (Transformers or SSMs) and training best practices.
• Depth in at least one area: (a) discrete speech tokens/temporal compression, (b) modality alignment to LLMs via adapters, or (c) post‑training/instruction tuning for speech tasks.
• Strong experimentation habits: clean code, ablations, reproducibility, and clear reporting.

Preferred Qualifications

• Experience with speech generation (neural codecs/vocoders) or hybrid text+speech decoding.
• Background in multilingual or code‑switching speech and domain adaptation.
• Hands‑on work evaluating safety, bias, hallucination, or spoofing risks in speech systems.
• Distributed training/serving (FSDP/DeepSpeed), and experience with ESPnet, SpeechBrain, or NVIDIA NeMo.

Tech Stack

• PyTorch, CUDA, torchaudio/librosa; experiment tracking (e.g., Weights & Biases).
• LLM backbones with lightweight adapters; neural audio codecs and vocoders as needed.
• FastAPI/gRPC for services; ONNX/TensorRT and quantization for efficient inference.

Logistics

• Location: Redmond(Preferred) or Remote

What We Offer

• Competitive stipend and hands-on projects with measurable real-world impact.
• Mentorship from applied scientists and engineers; opportunities to publish and present.
• Access to modern GPU infrastructure and a supportive environment for fast, responsible experimentation.
• Flexible location and schedule options, subject to team needs.