Bernd Huber

Bernd Huber

Research scientist · New York
PhD Computer Science, Harvard University

I work on audio, language models, and learning from human preference. My research includes reward modeling and post-training for AI agents, personalized generation, and models of speech and conversation.

Selected work

All publications

Reward modeling for music agents

At Spotify, I combined reward modeling with direct preference optimization (DPO) to train music agents from listening feedback. The approach improves tool-use plans using feedback on the playlists they produce. Production A/B tests showed 70% fewer erroneous tool calls and 4% more listening time.

Training loop: generate orchestration plans, score with a reward model, sample preference pairs, and fine-tune with DPO.
Preference optimization from listening feedback.

Embedding-to-Prefix

Embedding-to-Prefix (E2P) maps existing user embeddings to a single learned soft prefix token, personalizing a frozen LLM without updating its weights. I developed the method at Spotify; it was used for long-term personalization in GLIDE, a generative podcast recommender.

A user embedding is projected into one prefix token, which conditions a frozen transformer to generate personalized text.
Personalization through a learned prefix; the language model stays frozen.

Models of conversational speech

I developed SpecialTime and Meeter to analyze spoken conversation. SpecialTime provides real-time feedback on dialogue acts in parent–child speech. Meeter uses transcription, sentence segmentation, and classification to measure information sharing and shared understanding in group discussions. My speech research also includes real-time emotion recognition on mobile devices.

Meeter processes speech into a transcript and segmented sentences, then extracts word counts, information sharing, shared understanding, and prosodic features.
Meeter’s speech-processing pipeline, from the paper.

Other work

I also co-authored work on data selection for continual adaptation of generative recommenders. At Adobe Research, I developed B-Script, a system for transcript-based B-roll recommendations (US patent).

PhD thesis