语音识别是怎么听的How Speech Recognition Works
How Speech Recognition Works
把话转成文字的主流做法：弱监督大规模训练，一个模型听多种口音。The mainstream way to turn speech into text: large-scale weak supervision, one model for many accents.
Whisper（OpenAI，2022）用 68 万小时网络音视频做弱监督训练。一套模型覆盖多语言与多口音。
它把语音识别变成「翻译式」任务：音频进编码器，文字像解码输出一样逐词生成。
弱监督的代价：训练数据没有人工校对，转写在专业词、口音边缘上会出错。
评测口径是词错率（WER）：错词数除以总词数，越低越好。
Whisper (OpenAI, 2022) trains on 680,000 hours of web audio-video with weak supervision; one model covers many languages and accents.
It turns recognition into a translation-style task: audio goes into an encoder, text comes out as decoded output word by word.
The cost of weak supervision: training data has no human proofreading, so transcripts fail on specialist terms and accent edges.
The evaluation metric is word error rate (WER): wrong words divided by total words — lower is better.
语音功能的错误率直接决定后续理解质量。知道它是「统计转写」而不是「听懂」，才知道要在哪一环兜底。The error rate of the speech stage directly determines the quality of everything after it. Knowing it is statistical transcription rather than understanding tells you where to add a safety net.
回到名词地图看第4站的其他 35 个词 →Back to the map: the other 35 terms in station 4 →
- Whisper 论文《Robust Speech Recognition via Large-Scale Weak Supervision》（arXiv）Robust Speech Recognition via Large-Scale Weak Supervision (arXiv 2212.04356) ↗