Will Rice

ML engineer, seven years in speech and audio.

I work on real-time lip sync — audio-driven talking heads, and the inference work that gets them running in real time: 2-step distillation, distilled VAE, FP8 quantization, TensorRT. Before that, production TTS and STT serving on multi-GPU infrastructure, streaming pipelines with endpointing, voice activity detection and CTC decoding, and on-device ASR and NLU under embedded constraints.

Selected work

TFHubert / TFWav2Vec2	TensorFlow implementations of Meta's self-supervised speech models, in huggingface/transformers
vadkit	Multi-provider voice activity detection — FireRedVAD, Silero, FSMN-VAD, WebRTC — behind one streaming API for endpointing
pe-av-syncnet	SyncNet on Meta's Perception Encoder audio-visual representations, for lip-sync evaluation
whisper-rl	Reinforcement learning on Whisper, past what supervised fine-tuning reaches
denoisers	PyTorch waveform denoisers for speech enhancement
spokestack-python	Python library for embedded speech: on-device wake word, ASR and natural language understanding
ai-agent-security-2026	Kaggle gold — red-teaming LLM agents for multi-step tool-misuse attacks

Consulting

Available for consulting on speech, ASR/TTS and inference optimization.

wrice20@gmail.com · LinkedIn · Kaggle