whisper-base-fine_tuned-ru
| Property | Value |
|---|---|
| Parameter Count | 72.6M |
| License | Apache 2.0 |
| Base Model | openai/whisper-base |
| WER Score | 41.22% |
What is whisper-base-fine_tuned-ru?
whisper-base-fine_tuned-ru is a specialized automatic speech recognition (ASR) model optimized for the Russian language. Built upon OpenAI's Whisper base architecture, this model has been fine-tuned using the Mozilla Common Voice 11.0 dataset to enhance its performance specifically for Russian speech transcription.
Implementation Details
The model utilizes a transformer-based architecture with 72.6M parameters and F32 tensor precision. Fine-tuning was conducted using PyTorch with Native AMP mixed precision training, implementing an Adam optimizer with carefully tuned hyperparameters (β1=0.9, β2=0.999, ε=1e-08).
- Training batch size: 16 (4 base × 4 gradient accumulation steps)
- Learning rate: 1e-06 with linear scheduler
- Training steps: 20,000 with 250 warmup steps
- Achieved final validation loss: 0.4553
Core Capabilities
- Russian speech recognition with 41.22% WER
- Optimized for Russian language audio transcription
- Compatible with standard Whisper inference pipelines
- Supports TensorBoard logging for monitoring
Frequently Asked Questions
Q: What makes this model unique?
This model is specifically optimized for Russian language ASR, showing significant improvement through extensive fine-tuning on Russian speech data, starting from a 71.67% WER and improving to 41.22% WER through training.
Q: What are the recommended use cases?
The model is best suited for Russian speech transcription tasks, particularly in applications requiring automatic subtitling, transcription services, or voice command systems for Russian-speaking users.
