diarizers-community/DiarizationLM-Gemma-4-E4B-v1

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 6, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

DiarizationLM-Gemma-4-E4B-v1 is a 7.9 billion parameter DiarizationLM built on Google's Gemma 4 E4B foundation model, fine-tuned for post-processing automatic speech recognition (ASR) and speaker diarization outputs. It is trained with Locality-Preserving Oracle Supervision across four canonical speaker diarization benchmarks, including Fisher English, Callhome American English, ICSI Meeting Corpus, and AMI Meeting Corpus. This model excels at correcting lexical turn boundaries and backchannels while preserving acoustic speaker anchors, achieving statistically significant improvements in WDER and cpWER across all benchmarks compared to previous models, despite having half the parameter count.

Loading preview...

Overview

DiarizationLM-Gemma-4-E4B-v1 is a specialized Large Language Model (LLM) designed to enhance the accuracy of automatic speech recognition (ASR) and speaker diarization. Built upon Google's Gemma 4 E4B foundation model, this 7.9 billion parameter model is uniquely fine-tuned using Locality-Preserving Oracle Supervision.

Key Capabilities

  • Advanced Post-Processing: Corrects and optimizes ASR and speaker diarization outputs, specifically addressing lexical turn boundaries and backchannels (1-5 words).
  • Speaker Anchor Preservation: Maintains acoustic speaker anchors across long monologues (6+ words), a common challenge in multi-speaker scenarios.
  • Broad Benchmark Training: Fine-tuned across all four canonical speaker diarization benchmarks: Fisher English, Callhome American English, ICSI Meeting Corpus, and AMI Meeting Corpus.
  • Superior Performance: Achieves statistically significant (p < 0.0001) improvements in Word Diarization Error Rate (WDER) and concatenated Purity-Weighted Error Rate (cpWER) across all training benchmarks, outperforming larger 8B parameter models in some cases.
  • Efficient Architecture: Delivers strong performance with a 4B dense parameter count, making it more efficient than previous 8B parameter DiarizationLM models.

Good For

  • Improving Diarization Accuracy: Ideal for applications requiring highly accurate speaker segmentation and attribution in conversational audio.
  • Multi-Speaker Environments: Particularly effective in complex scenarios like telephone conversations and multi-person meetings (2-9 speakers).
  • Integration with Open-Source Pipelines: Seamlessly integrates with the diarizers open-source ASR + Diarization + DiarizationLM pipeline for end-to-end audio processing.
  • Resource-Efficient Deployment: Available in a 4-bit K-quant Medium (Q4_K_M) GGUF format, suitable for CPU deployment via llama.cpp or Ollama.

Finetunes
1 models