Qwen3-TTS for Swiss German: Dialect Adaptation via Talker Layer Fine-Tuning
How we adapted Qwen3-TTS to Swiss German by training only the last 8 talker layers and creating a dedicated speaker per dialect — with audio examples from ou...
Blog
Field notes, thoughts, experiments, and shipping lessons from production ASR and NLP work and other cool stuff at work.
2026
My Personal AI: Living with OpenClaw
From Telegram pings to automated calorie tracking and agentic coding-how OpenClaw changed my relationship with AI.
2025
The Case of the Diverging Loss: When ‘Best Loss’ != ‘Best Model’
Exploring the counter-intuitive relationship between training loss and Word Error Rate (WER) during Whisper fine-tuning.
Visualizing Whisper: Mapping Cross-Attention to Time
How decoder cross-attention weights can be used to create high-precision token-to-time alignments without external models.
Solving the game “Overcooked” with Reinforcement Learning
Solving the Overcooked Game with PPO.
Numerai – Top 10 Participant and Numerai Master
Competing at the top of the Numerai tournament with live-deployed ML models
SwissText 2025 - Whisper Finetuning
Presentation about Fine-tuning Whisper for Swiss-German at SwissText 2025
Imaging Oxygen Dips in Mouse Cortex: Semi-NMF for Signal Extraction
Science Publication: We segment and extract transient oxygen ‘dips’ from bioluminescence recordings using semi-NMF.