Google's DiarizationLM-Gemma fixes speaker attribution, hits 2.99% WER on Fisher
lmoroney · x · 2026-10-06
Google's speaker team released DiarizationLM-Gemma-4-E4B-v1 on Hugging Face, a Gemma 4 E4B fine-tune that post-processes finished transcripts to fix speaker attribution (e.g., mispinned "mm-hmm" backchannels). Training lets it move 1-5 word backchannels while treating utterances of 6+ words as anchors, keeping long monologues attributed correctly.
- Improves speaker attribution error on all four test sets, including 4-9 person meetings from ICSI and AMI
- Reaches 2.99% error on Fisher phone calls vs 3.28% for the earlier 8B version
- 4-bit GGUF is 5.3 GB, runnable on desktop
Suggested as a post-processing step; measure on a few hours of your own labeled meetings before relying on it.
More from coding & agent
- Zeroization can make things worse: how wiping secrets creates more copies — jedisct1 · 2026-10-06
- Indie dev launches agent-first creator marketing platform Clipatra, pays out $3,000+ — tibo_maker · 2026-10-06
- TensorFold bonds dual Thunderbolt 5 links for 83% throughput boost on Apple Silicon — AIFlow_ML · 2026-10-06
- A 45-minute visual tour of graph theory, taught through the author's hometown — TivadarDanka · 2026-10-06
- Open-weights Kolibri-1 plays Breakout with no fine-tuning at ~25ms per move — Nils_Reimers · 2026-10-06
- Gatana adds centralized skills sync to MCP Gateway across all agents — Gatana_Official · 2026-10-06