ASR-MMS-SALT Speech Recognition Model
| Property | Value |
|---|---|
| License | CC-BY-NC-4.0 |
| Base Model | facebook/mms-1b-all |
| Supported Languages | 6 (Luganda, Acholi, Lugbara, Ateso, Runyankole, English) |
| Downloads | 36,557 |
What is asr-mms-salt?
ASR-MMS-SALT is a specialized speech recognition model designed specifically for Ugandan languages. Built upon Facebook's MMS-1B-ALL architecture, this model represents a significant advancement in African language processing, supporting six languages with unique code-switching capabilities.
Implementation Details
The model utilizes the wav2vec2 architecture and implements language-specific adapters for each supported language. It's built using PyTorch and includes both monolingual and code-switching variants for each language, making it highly versatile for real-world applications.
- Dual adapter system for each language: pure language and English code-switching
- Built on transformer architecture with wav2vec2 implementation
- Trained on the SALT dataset specifically for Ugandan languages
- Outputs unpunctuated, lowercase text
Core Capabilities
- Accurate speech recognition in 6 Ugandan languages
- Support for English code-switching scenarios
- Efficient processing with CUDA support
- Easy integration with the Transformers library
- Specialized handling of regional accents and dialects
Frequently Asked Questions
Q: What makes this model unique?
The model's distinctive feature is its specialized support for Ugandan languages with dedicated adapters for both pure language and code-switching scenarios, making it particularly valuable for real-world applications in East Africa.
Q: What are the recommended use cases?
This model is ideal for applications requiring speech recognition in Ugandan languages, particularly in scenarios where speakers might switch between local languages and English. It's suitable for transcription services, voice assistants, and language documentation projects in East Africa.
