Whisper Medium.en
| Property | Value |
|---|---|
| Parameter Count | 764M parameters |
| License | Apache 2.0 |
| Paper | Robust Speech Recognition via Large-Scale Weak Supervision |
| Test WER (LibriSpeech Clean) | 4.12% |
What is whisper-medium.en?
Whisper-medium.en is a powerful English-specific automatic speech recognition (ASR) model developed by OpenAI. It's based on a Transformer encoder-decoder architecture and has been trained on 680,000 hours of labeled speech data. This particular variant is optimized specifically for English language transcription, making it more efficient for English-only use cases compared to its multilingual counterparts.
Implementation Details
The model employs a sequence-to-sequence architecture utilizing the Transformer framework. With 764M parameters, it sits in the middle of OpenAI's Whisper model range, offering a good balance between accuracy and computational requirements.
- Transformer-based encoder-decoder architecture
- Trained on 438,000 hours of English audio data
- Supports long-form transcription through 30-second chunking
- Includes built-in support for timestamp prediction
Core Capabilities
- High-accuracy English speech transcription
- Robust performance across different accents and background noise
- Batch processing support for efficient transcription
- Zero-shot generalization to various domains
- Support for timestamp generation
Frequently Asked Questions
Q: What makes this model unique?
The model's unique strength lies in its robust generalization capabilities without requiring fine-tuning, achieved through extensive training on 680k hours of diverse audio data. It demonstrates exceptional performance on English ASR tasks with a 4.12% WER on LibriSpeech clean test set.
Q: What are the recommended use cases?
The model is ideal for English speech recognition tasks, particularly in scenarios requiring high accuracy and robustness to different accents and background noise. It's well-suited for transcription services, accessibility tools, and research applications, though real-time transcription would require additional optimization.
