nb-wav2vec2-300m-nynorsk
| Property | Value |
|---|---|
| Parameter Count | 315M |
| License | Apache 2.0 |
| Paper | arXiv:2307.01672 |
| WER Score | 12.22% (with KenLM) |
What is nb-wav2vec2-300m-nynorsk?
nb-wav2vec2-300m-nynorsk is a specialized automatic speech recognition (ASR) model designed for Norwegian Nynorsk language. Built on the VoxRex feature extractor from the National Library of Sweden, this model represents a significant advancement in Norwegian language processing technology. The model achieves a Word Error Rate (WER) of 12.22% with a 5-gram KenLM language model integration.
Implementation Details
The model was developed during the Hugging Face Robust Speech Event and leverages the Norwegian Parliamentary Speech Corpus (NPSC). It utilizes a wav2vec2 architecture with 315M parameters and includes specific optimizations for Norwegian Nynorsk recognition.
- Fine-tuned on the NbAiLab/NPSC dataset
- Implements gradient checkpointing for efficient training
- Features customizable mask time and feature probabilities
- Includes freeze feature encoder optimization
Core Capabilities
- Specialized Norwegian Nynorsk speech recognition
- Achieves 12.22% WER with language model (15.37% without)
- Character Error Rate (CER) of 4.19% with language model
- Supports audio inputs between 0.5 and 30 seconds
Frequently Asked Questions
Q: What makes this model unique?
This model is specifically optimized for Norwegian Nynorsk, one of Norway's official written languages, and represents part of a comprehensive effort to improve Norwegian ASR technology. Its integration with a 5-gram KenLM language model significantly improves recognition accuracy.
Q: What are the recommended use cases?
The model is ideal for Norwegian Nynorsk speech transcription tasks, particularly in parliamentary and formal speech contexts. It's suitable for applications requiring moderate to high-accuracy ASR for Norwegian Nynorsk, with audio durations between 0.5 and 30 seconds.
