The most hidden error in meeting transcripts is rarely a wrong word — it is a wrong speaker attribution: the manager's "yes, let's do it" lands under the questioner's name, and every downstream summary inherits the wrong ownership. This month Google quietly published a small model on Hugging Face, DiarizationLM-Gemma-4-E4B-v1, built to fix exactly that. The bet is not fuller transcription but more correct attribution of who said what.
It sits at the tail of the pipeline
The roughly 4B-parameter model is a Gemma fine-tune. Its input is not raw audio but the text-plus-speaker-labels already produced by upstream ASR and speaker diarization. It reads the conversational context, spots turn assignments that make no linguistic sense, and corrects them. The model card is equally clear about the limits: it depends on upstream output, and if the original transcription already dropped content, no post-processor can conjure the acoustic evidence back.
Training idea: keep acoustic anchors across long monologues
According to the official model card, earlier DiarizationLM models (such as DiarizationLM-8b-Fisher-v2) were trained mostly on two-speaker telephone data, and naive multi-domain SFT suffers from speaker-identity drift in meetings with four or more speakers during long monologues. The new release trains on locality-preserving oracle targets: the model learns to correct lexical turn boundaries and backchannels of 1-5 words while preserving acoustic speaker anchors across monologues of 6 words or more.
Read the numbers separately
The official evaluation covers four benchmarks, and word-level diarization error rate (WDER — the share of words assigned to the wrong speaker) drops on all of them, but by very different margins: Fisher falls from 5.32% to 2.99%, nearly halved; the meeting corpora move far less, ICSI from 14.70% to 14.10% and AMI from 15.68% to 14.89%. Independent coverage likewise notes that two of the four gains are tiny. Post-processing has real potential, but one strong dataset cannot be extrapolated into a universal claim.
Boundaries and judgment
The model card states plainly that this is a research release, not an officially supported Google product. Two takeaways. First, this is pragmatic engineering: instead of an end-to-end speech model, an LLM is inserted into an existing pipeline as a context-based proofreader, and a 4B footprint keeps it cheap. Second, the risk hides inside the strength: correcting labels by linguistic context means a fix that reads more smoothly can paper over what was actually said, so consequential meetings should always keep the original audio as the checkable ground truth. Teams building meeting-minutes products should A/B this inside their pipeline, with "attribution accuracy" — not "smoother transcripts" — as the acceptance bar.
So the next race in speech AI is not transcribing more, but attributing correctly. Sources: official model card and technical walkthrough.