14MB印地英语语音识别模型实现29%的pWER，生词率2.3/100
来源标题：Fourteen MB Hinglish STT pWER 29, stary words 2.3/100
一个仅14MB的印地英语（Hinglish）语音识别模型达到29%的pWER，生词率为每100词2.3个。该结果体现了模型在极小体积下的识别表现。
以下正文同步自 Hacker News 最新，版权归原站所有，已转换为易读排版。
pWER, lower is better. Single run on frozen Hinglish and Hindi test splits, greedy decoding, measured from the exported 4-bit file. The other models were not tuned for these splits. Methods and caveats are on the model card.
pWER on frozen test splits, single run. Lower is better.
Hinglish error on hn_gold, same audio, same scorer. Single run on our test splits; the other models were not tuned for them.
Ready to hear your users, in their words?
Download the weights, run them on 50 of your own clips, and let your pWER decide. Free for non-commercial use under CC BY-NC-SA 4.0; need commercial rights? Get in touch.
Public weights, weak spots named, no sign-up.
正文由 探How 从来源站点 RSS 同步，内容未经改写；遇到排版缺失或需要图片、视频时请以原文为准。