DVoice Kabyle ASR Model
| Property | Value |
|---|---|
| License | Apache 2.0 |
| Framework | SpeechBrain, PyTorch |
| Architecture | wav2vec 2.0 + CTC/Attention |
| Test Performance | WER: 24.80%, CER: 6.55% |
What is dvoice-kabyle?
DVoice-kabyle is an automatic speech recognition (ASR) model specifically designed for the Kabyle language, developed by AIOX Labs. It leverages the powerful wav2vec 2.0 architecture combined with CTC/Attention mechanisms to provide accurate speech-to-text conversion for Kabyle speakers. The model was trained on the CommonVoice dataset and represents a significant step forward in low-resource language technology.
Implementation Details
The model implements a two-block architecture consisting of a unigram tokenizer for subword unit transformation and an acoustic model based on a fine-tuned wav2vec 2.0 (facebook/wav2vec2-large-xlsr-53). The system processes 16kHz single-channel audio and includes automatic audio normalization capabilities.
- Pre-trained wav2vec 2.0 base model with additional DNN layers
- CTC greedy decoder for final transcription
- Automatic audio normalization and resampling
- SpeechBrain framework integration
Core Capabilities
- Accurate Kabyle speech recognition with 24.80% WER on test data
- Real-time audio transcription support
- GPU inference compatibility
- Automatic audio preprocessing
Frequently Asked Questions
Q: What makes this model unique?
This model is specifically designed for the Kabyle language, which is typically considered a low-resource language. It's part of the larger DVoice initiative aimed at providing voice technologies for African languages, making it a crucial tool for Kabyle language technology development.
Q: What are the recommended use cases?
The model is ideal for Kabyle speech transcription tasks, including content creation, accessibility services, and language preservation efforts. It's particularly suitable for applications requiring 16kHz audio processing and can be deployed in both CPU and GPU environments.
