USM（Universal Speech Model）

通用语音模型 (USM) 是一系列最先进的语音模型，其 2B 参数经过 1200 万小时的语音和 280 亿条文本句子的训练，涵盖 300 多种语言。 USM 用于 YouTube（例如，隐藏式字幕），可以对广泛使用的语言（如英语和普通话）以及旁遮普语、阿萨姆语、桑塔利语、巴厘岛语、绍纳语、马达加斯加语等语言执行自动语音识别 (ASR)， Luganda、Luo、Bambara、Soga、Maninka、Khosa、Akan、Lingala、Chichewa、Nkore、Nzema 等等。 其中一些语言只有不到 2000 万人使用，因此很难找到必要的训练数据。

Universal Speech Model (USM) is a series of state-of-the-art speech models. Its 2-billion-parameter variant is trained on 12 million hours of speech data and 28 billion text sentences, covering over 300 languages. USM is deployed in YouTube applications such as closed captions, and supports automatic speech recognition (ASR) for a wide range of languages, including widely used ones like English and Mandarin, as well as less prevalent languages such as Punjabi, Assamese, Santali, Balinese, Shona, Malagasy, Luganda, Luo, Bambara, Soga, Maninka, Khosa, Akan, Lingala, Chichewa, Nkore, Nzema, and many others. Some of these languages have fewer than 20 million speakers, making it particularly challenging to obtain the necessary training data.

搜集汇总

背景与挑战

背景概述

USM（通用语音模型）是一个包含2B参数的语音模型系列，基于1200万小时语音和280亿条文本句子训练，覆盖300多种语言。该模型可用于自动语音识别，如YouTube隐藏式字幕，支持英语、普通话以及多种低资源语言。