Summary of Efficient Compression Of Multitask Multilingual Speech Models, by Thomas Palmeira Ferraz
Efficient Compression of Multitask Multilingual Speech Models
by Thomas Palmeira Ferraz
First submitted to arxiv on: 2 May 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); Sound (cs.SD); Audio and Speech Processing (eess.AS)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The Whisper model, a multitask and multilingual speech model covering 99 languages, demonstrates commendable automatic speech recognition (ASR) results in a subset of its covered languages. However, it still underperforms on a non-negligible number of under-represented languages, particularly in smaller model versions. This paper examines the limitations of Whisper, revealing speaker-related and model-related bias. Despite these limitations, the authors propose DistilWhisper, an approach that bridges the performance gap in ASR for low-resource languages while retaining multitask and multilingual capabilities. The dual approach involves lightweight modular ASR fine-tuning using language-specific experts and knowledge distillation from whisper-large-v2. Results show that this approach outperforms standard fine-tuning or LoRA adapters, boosting performance in targeted languages for both in- and out-of-domain test sets while introducing only a negligible parameter overhead at inference. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Whisper is a special kind of AI model that can understand and recognize speech in many different languages. It’s really good at recognizing speech in some languages, but not as good in others – especially languages that don’t have as much data or information available. This paper looks at why Whisper isn’t perfect and tries to find ways to make it better. The solution involves making a new model called DistilWhisper that can learn from other models and improve its ability to recognize speech in different languages. |
Keywords
» Artificial intelligence » Boosting » Fine tuning » Inference » Knowledge distillation » Lora
