About The Role

Boson AI builds real-time speech and multimodal AI that makes talking to AI feel like talking to people. Our Higgs models listen, speak, and see. Higgs Realtime is a speech-to-speech API with tool calling and support for 100+ languages, and it is drop-in compatible with the OpenAI Realtime API. Higgs TTS 3 is an expressive, open-weight text-to-speech model with zero-shot voice cloning. Higgs Avatar turns a single image into a streaming, talking video, and Higgs VoiceOver brings directable AI voice-over into video editors. We are a team of 40+ researchers and engineers, and we train our models on our own cluster of 900 GPUs.

Eligibility: this internship is only open to students enrolled in the Master of Science in Applied Computing (MScAC) program at the University of Toronto. Applications from anyone who is not an MScAC student at the University of Toronto will be automatically declined.

We are hiring MScAC students for the program's 8-month applied research internship, starting in May 2027; an earlier start is possible. You will join one of the projects below and work with a Boson AI mentor on problems that feed directly into our models and products. The internship is based in our Toronto office on a hybrid schedule.

Research Projects

- Audio data extraction and annotation. Our speech models learn from very large amounts of audio in 100+ languages. You will build pipelines that collect, segment, transcribe and annotate audio with as little manual engineering as possible, using unsupervised and semi-supervised methods, and you will close the loop so that annotation and model training improve each other (see NVIDIA Granary, arXiv:2505.13404). This suits you if you enjoy scripting, experiments and large datasets,



and have a working knowledge of statistics and machine learning.
- Audio evaluation and benchmarking. Positive speech is more than correct pronunciation. It has to fit the context, emotion and style, sound natural in a conversation, and match the voice it clones. You will design benchmarks and automatic metrics for expressive, multilingual and conversational speech, for voice cloning, and for speech recognition under the conditions real voice agents face (see our EmergentTTS-Eval, arXiv:2505.23009, and Back to Basics: Revisiting ASR in the Age of Voice Agents, arXiv:2603.25727), and use them both to evaluate our models and as signals for training them. This suits you if you like experimentation, have a good ear, ideally speak more than one language, and are comfortable automating evaluation in Python.
- Video generation and computer vision. Higgs Avatar turns a single image into a streaming, talking video. You will work on the data pipeline behind it (collection, multi-stage filtering and synthesis) and on the models: audio-driven video generation with accurate lip sync and natural expressions, distilled diffusion models that run in real time, and computer vision that lets our models see and react to the person they are talking to (see Wan-S2V, arXiv:2508.18621, and LTX-Video, arXiv:2501.00103). This suits you if you have a background in image or video generation (diffusion models, VAEs,



GANs, NeRFs) or in computer vision, and are comfortable with large-scale data and model optimization.
- Efficient training. You will make training on our 900-GPU cluster faster and more robust: numerical formats (FP8, FP4, BF16), fused CUDA or Triton kernels for architectures that go beyond standard Transformers, overlapping communication with computation, faster checkpointing and recovery from failures, newer optimizers such as Muon, and stable reinforcement-learning post-training when the training and rollout systems differ (see our Let the Batch Decide, arXiv:2605.12380). This suits you if you enjoy systems work and optimization, like writing CUDA or Triton, and want experience training across hundreds of GPUs. Strong coding skills are required.
- Serving and real-time inference. Higgs Realtime responds in under a second, and Higgs Avatar streams live video into calls. You will make our streaming speech and video models faster and cheaper to serve, on our own GPUs and in the cloud, using SGLang or vLLM: 8-bit and 4-bit quantization, speculative decoding, kernel and memory-layout optimization, KV-cache management across many concurrent users, and scaling capacity with demand (see our Alignment Drift in Single-Model Speculative Decoding for ASR, arXiv:2608.12703, and TARQ, arXiv:2605.27808). This suits you if you enjoy systems work, have some experience with GPU hardware and CUDA or Triton, and want to work with large numbers of GPUs. Strong coding skills are required.
- Agents: conversational design, theory of mind and reinforcement learning. Our voice agents handle customer support, sales and training calls for businesses. They must foll

📌 Internship - MScAC 2027 (Toronto)
🏢 Boson
📍 Toronto