DiarizationLM-8b-Fisher-v1 is an 8 billion parameter DiarizationLM model developed by Google, fine-tuned on the Fisher corpus for speaker diarization post-processing. Built upon the unsloth/llama-3-8b-bnb-4bit foundation model, it specializes in refining speaker attribution in audio transcripts. This model improves Word Diarization Error Rate (WDER) on the Fisher testing set, offering enhanced accuracy for diarization tasks.
Loading preview...
DiarizationLM-8b-Fisher-v1 Overview
This model, developed by Google, is an 8 billion parameter DiarizationLM fine-tuned specifically for speaker diarization post-processing. It is built on the unsloth/llama-3-8b-bnb-4bit foundation model and was trained using a LoRA adapter on the training subset of the Fisher corpus. The training involved approximately 8 epochs over 25,400 steps, utilizing a mixed data flavor combining hyp2ora and deg2ref prompt-completion pairs.
Key Capabilities
- Speaker Diarization Post-Processing: Designed to refine speaker attribution in transcribed audio.
- Performance Improvement: Demonstrates an improved Word Diarization Error Rate (WDER) of 4.40% on the Fisher testing set, compared to a baseline of 5.32%.
- Context Handling: Supports a maximal prompt length of 6000 characters and a maximal sequence length of 4096 tokens.
Good For
- Enhancing Diarization Accuracy: Ideal for applications requiring more precise speaker identification in audio transcripts.
- Research and Development: Useful for researchers working on speaker diarization and large language model applications in audio processing.
Note: This model is considered outdated; users are advised to use google/DiarizationLM-8b-Fisher-v2 for the latest version.
Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.