whisper-tiny-russian-dysarthria

Maintained By
qymyz

whisper-tiny-russian-dysarthria

PropertyValue
Base ModelOpenAI Whisper-tiny
TaskRussian Speech Recognition
WER Score9.1029%
Authorqymyz
Model URLHuggingFace

What is whisper-tiny-russian-dysarthria?

This model is a specialized fine-tuned version of OpenAI's Whisper-tiny model, specifically optimized for processing Russian dysarthric speech. Dysarthria is a motor speech disorder, and this model aims to improve speech recognition accuracy for affected individuals speaking Russian.

Implementation Details

The model was trained using carefully selected hyperparameters, including an Adam optimizer with a learning rate of 1e-05 and linear scheduling. Training was conducted over 3000 steps with a batch size of 16, incorporating native AMP (Automated Mixed Precision) training for optimal performance.

  • Learning rate: 1e-05 with linear scheduler and 500 warmup steps
  • Batch size: 16 for both training and evaluation
  • Training duration: 3000 steps across 30 epochs
  • Final validation loss: 0.2158

Core Capabilities

  • Specialized in Russian speech recognition
  • Optimized for dysarthric speech patterns
  • Achieves 9.1% Word Error Rate (WER)
  • Supports real-time transcription with efficient processing

Frequently Asked Questions

Q: What makes this model unique?

This model is specifically optimized for Russian dysarthric speech, making it particularly valuable for applications involving speakers with speech motor disorders. The low WER of 9.1% demonstrates its effectiveness in this specialized domain.

Q: What are the recommended use cases?

The model is ideal for applications requiring Russian speech recognition for individuals with dysarthria, such as assistive technologies, medical applications, and accessibility tools. It's particularly suitable for scenarios where accurate transcription of impaired speech is crucial.

🍰 Interesting in building your own agents?
PromptLayer provides Huggingface integration tools to manage and monitor prompts with your whole team. Get started here.