| Property | Value |
|---|---|
| Model Size | 2 Billion parameters |
| Type | Speech Recognition (ASR) |
| Language | Dutch/Flemish |
| WER | 3.93% |
| CER | 1.22% |
| Author | FremyCompany |
| Model URL | Hugging Face |
What is xls-r-2b-nl-v2_lm-5gram-os2_hunspell?
This is a sophisticated Dutch speech recognition model based on Facebook's XLS-R architecture, enhanced with a 5-gram language model and Hunspell-based typo correction. The model processes 16kHz audio input and employs a Wav2Vec2ForCTC decoder with 48 letters to generate letter-transcription probabilities per frame.
Implementation Details
The model combines three powerful components: a base XLS-R model, a language model, and a spelling correction system. It was trained through a carefully designed process involving multiple epochs on both Common Voice 8.0 and CGN datasets.
- Initialized with Facebook's 2B parameter XLS-R model
- Uses pyctcdecode-based beam search with 5-gram language model
- Incorporates Hunspell for intelligent typo correction
- Optimized for 12s chunks with 2s stride
Core Capabilities
- High-accuracy Dutch speech recognition with 3.93% WER
- Advanced typo correction using Hunspell dictionary
- Handles both Dutch and Flemish spoken variants
- Optimized for production environments
- Transcribes speech without punctuation
Frequently Asked Questions
Q: What makes this model unique?
The model's uniqueness lies in its combination of a powerful XLS-R base model with a sophisticated 5-gram language model and Hunspell-based spelling correction, specifically optimized for Dutch language recognition.
Q: What are the recommended use cases?
This model is ideal for transcribing Dutch or Flemish spoken content to text. It's particularly suitable for applications requiring high accuracy in speech recognition, though users should note it doesn't include punctuation in its output.