Improvements from Model Error Analysis
#1
by aisyahhrazak - opened
- To improve on numbers and silent sound robustness
- add more data on numbers
- add noisy data with silent (model tend to hallucinates when silent)
For add noisy data with silent, I think because I filtered the dataset based on words distance for an example https://github.com/malaysia-ai/dataset/blob/main/streaming-stt/science-streaming.ipynb,
if i > 0 and (d[i]['start'] - d[i - 1]['end']) > 3: