VoiceArena:更多数据胜过更大模型,微调 Whisper 实现孟加拉语
For ASR (automatic speech recognition) on weak languages, more training data ten…
For ASR (automatic speech recognition) on weak languages, more training data tends to beat a bigger model.
对于弱资源语言(weak languages)的自动语音识别(ASR),更多的训练数据往往比更大的模型效果更好。
@voicearena_ai’s Bengali result for Monsoon, its ASR corpus, is shows it.
@voicearena_ai 针对 Monsoon(其 ASR 语料库)提供的孟加拉语结果证实了这一点。
Whisper Medium fine-tuned on Monsoon went from 85.27% to 7.65% LLM word error rate on Bengali FLEURS.
在 Monsoon 上微调后的 Whisper Medium,其在孟加拉语 FLEURS 上的 LLM 词错误率从 85.27% 降至 7.65%。
Medium is the 769M-parameter Whisper. Small enough to serve cheaply, and in many settings small enough to run close to the user.
Medium 是拥有 7.69 亿参数的 Whisper 模型。它足够小巧,可以低成本部署,并且在许多场景下足以在用户端本地运行。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力