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, specifically designed for post-processing and optimizing automatic speech recognition (ASR) and speaker diarization outputs. It is fine-tuned with Locality-Preserving Oracle Supervision across four canonical speaker diarization benchmarks, enabling it to correct lexical turn boundaries and backchannels while preserving acoustic speaker anchors in multi-speaker conversations. This model achieves statistically significant improvements in Word Diarization Error Rate (WDER) and Character-Per-Word Error Rate (cpWER) across all benchmarks, even with half the parameter count of previous models.

Loading preview...

Overview

DiarizationLM-Gemma-4-E4B-v1 is a specialized Large Language Model (LLM) from diarizers-community, built upon Google's Gemma 4 E4B foundation model (4 billion dense parameters). Its primary function is to enhance and correct outputs from automatic speech recognition (ASR) and speaker diarization systems. Unlike general-purpose LLMs, this model is uniquely fine-tuned using Locality-Preserving Oracle Supervision across a diverse set of four canonical speaker diarization benchmarks: Fisher English, Callhome American English, ICSI Meeting Corpus, and AMI Meeting Corpus.

Key Capabilities

  • Advanced Diarization Post-Processing: Corrects lexical turn boundaries and backchannels (1-5 words) while maintaining speaker identity across longer monologues (6+ words).
  • Multi-Speaker Robustness: Specifically trained to avoid speaker identity drift in complex, multi-speaker meeting scenarios (4+ speakers), a common issue with models trained on simpler 2-speaker data.
  • Superior Performance: Achieves statistically significant (p < 0.0001) improvements in both Word Diarization Error Rate (WDER) and Character-Per-Word Error Rate (cpWER) across all four benchmarks, outperforming earlier models with twice the parameter count.
  • Efficient Architecture: Despite its strong performance, it leverages a 4B parameter Gemma base, making it more efficient than larger alternatives like DiarizationLM-8b-Fisher-v2 (Llama 3 8B).

Good For

  • Improving ASR and Diarization Accuracy: Ideal for applications requiring highly accurate speaker attribution and transcription in conversational and meeting data.
  • Complex Multi-Speaker Scenarios: Particularly effective in environments with multiple speakers and varying conversation dynamics, such as conference calls, interviews, and academic meetings.
  • Research and Development: Provides a robust baseline and advanced capabilities for further research in speaker diarization and speech processing.
  • Resource-Efficient Deployment: The availability of a 4-bit GGUF quantized version (Q4_K_M) makes it suitable for deployment on devices with limited resources, including llama.cpp and Ollama.

Finetunes
1 models