1
00:00:00,000 --> 00:00:03,760
歡迎收聽聽書Podcast本大由壯菜工程師製作。

2
00:00:03,760 --> 00:00:11,840
那個現在想請你在腦海中稍微想像一下一個我們日常工作極度混亂的場景好了。

3
00:00:11,840 --> 00:00:13,160
嗯,多混亂。

4
00:00:13,160 --> 00:00:18,680
這是你正身處在一個可能有八個人同時講話的那種會議室裡面。

5
00:00:18,680 --> 00:00:20,080
哇,那真的很吵耶。

6
00:00:20,080 --> 00:00:21,520
對,超級吵。

7
00:00:21,520 --> 00:00:25,840
然後有人在為了那個專案的進度激烈爭論,

8
00:00:25,840 --> 00:00:28,880
然後旁邊可能還有人在小聲碎碎唸。

9
00:00:28,880 --> 00:00:31,680
重點是大家不但中英文夾雜

10
00:00:31,680 --> 00:00:34,640
還不斷互相搶話跟疊話

11
00:00:34,640 --> 00:00:36,040
這聽起來就是

12
00:00:36,040 --> 00:00:39,180
每天辦公室會發生的噩夢啊

13
00:00:39,180 --> 00:00:40,260
沒錯沒錯

14
00:00:40,260 --> 00:00:42,560
但這個時候如果你想像一下

15
00:00:42,560 --> 00:00:43,580
如果有一個助理

16
00:00:43,580 --> 00:00:45,600
能在你剛講完話的

17
00:00:45,600 --> 00:00:48,060
呃0.16秒以內

18
00:00:48,060 --> 00:00:50,340
立刻遞給你一份完美的逐字稿

19
00:00:50,340 --> 00:00:51,880
0.16秒

20
00:00:51,880 --> 00:00:53,500
這幾乎是同步了吧

21
00:00:53,500 --> 00:00:54,080
對

22
00:00:54,080 --> 00:00:56,280
而且這份逐字稿上面

23
00:00:56,280 --> 00:00:58,460
還清清楚楚的標記了

24
00:00:58,460 --> 00:01:00,620
刚刚哪一句话是谁说的

25
00:01:00,620 --> 00:01:03,920
甚至连两个人同时开口讲话的内容

26
00:01:03,920 --> 00:01:05,520
它都分得一清二楚

27
00:01:05,520 --> 00:01:08,960
这听起来真的很像某种科幻电影里

28
00:01:08,960 --> 00:01:10,300
才有的可怕科技对吧

29
00:01:10,300 --> 00:01:10,660
真的

30
00:01:10,660 --> 00:01:13,400
一般人听起来一定觉得这不可能啦

31
00:01:13,400 --> 00:01:15,540
但其实这已经不是科幻情节了

32
00:01:15,540 --> 00:01:17,240
它已经真实发生了

33
00:01:17,240 --> 00:01:19,940
就在2026年9月1日

34
00:01:19,940 --> 00:01:22,280
META的超级智慧实验室

35
00:01:22,280 --> 00:01:24,100
也就是那个MSL

36
00:01:24,100 --> 00:01:25,440
哦MSL

37
00:01:25,440 --> 00:01:25,740
对

38
00:01:25,740 --> 00:01:27,920
他们正式发布了最新的

39
00:01:27,920 --> 00:01:29,780
即时音讯感知模型

40
00:01:29,780 --> 00:01:32,040
叫做Muse Voice Transcribe

41
00:01:32,040 --> 00:01:33,640
Muse Voice Transcribe

42
00:01:33,640 --> 00:01:35,340
这名字听起来蛮厉害的

43
00:01:35,340 --> 00:01:36,320
这个技术啊

44
00:01:36,320 --> 00:01:38,860
它不仅做到了你刚刚描述的那种

45
00:01:38,860 --> 00:01:41,560
在极度混乱的场景下完美记录

46
00:01:41,560 --> 00:01:44,040
它甚至在多项基准测试里面

47
00:01:44,040 --> 00:01:47,480
直接击败了Google跟OpenAI的同级产品

48
00:01:47,480 --> 00:01:50,020
哇 直接打败两大巨头耶

49
00:01:50,020 --> 00:01:52,280
所以在这场为你量身打造的

50
00:01:52,280 --> 00:01:53,080
深度探跑中

51
00:01:53,080 --> 00:01:54,580
我们绝点不会只是你知道

52
00:01:54,580 --> 00:01:56,780
很无聊的去念那些声音的规格表

53
00:01:56,780 --> 00:01:58,180
对那太无聊了

54
00:01:58,180 --> 00:01:58,540
没错

55
00:01:58,540 --> 00:02:00,460
我们会带你彻底拆解

56
00:02:00,460 --> 00:02:02,880
这个打败众多科技巨头的核心技术

57
00:02:02,880 --> 00:02:04,460
去了解这个AI

58
00:02:04,460 --> 00:02:06,520
它是如何像我们人类大脑一样

59
00:02:06,520 --> 00:02:09,820
懂得在听话的时候去犹豫

60
00:02:09,820 --> 00:02:10,740
犹豫啊

61
00:02:10,740 --> 00:02:11,900
这个词用的很好

62
00:02:11,900 --> 00:02:12,380
对吧

63
00:02:12,380 --> 00:02:12,920
对吧

64
00:02:12,920 --> 00:02:16,360
然后这项技术将如何彻底颠覆你未来

65
00:02:16,360 --> 00:02:18,520
跟科技装饰互动的方式

66
00:02:18,520 --> 00:02:19,980
比如说你的Mac

67
00:02:19,980 --> 00:02:23,460
甚至是未来的那个AR智慧眼镜

68
00:02:23,460 --> 00:02:28,760
没错,不过其实要真正理解Muse Voice Transcribe到底有多强大

69
00:02:28,760 --> 00:02:30,300
我们必须先回头看看

70
00:02:30,300 --> 00:02:35,060
过去几年的语音辨识技术到底有多么的笨拙

71
00:02:35,060 --> 00:02:39,020
笨拙,所以过去的方法很卡就对了

72
00:02:39,020 --> 00:02:43,480
超级卡,我们得先看看过去的那种拼装车系统

73
00:02:43,480 --> 00:02:44,220
拼装车

74
00:02:44,220 --> 00:02:48,160
对,因为过去如果你想要完成一份会议的逐字稿

75
00:02:48,160 --> 00:02:51,220
它其实不是一个单一的人工智慧在运作

76
00:02:51,220 --> 00:02:55,060
它是需要三套完全独立的系统来接力完成

77
00:02:55,060 --> 00:02:56,980
呃哪三套

78
00:02:56,980 --> 00:02:58,720
首先是那个ASR

79
00:02:58,720 --> 00:03:00,880
也就是自动语音辨识系统

80
00:03:00,880 --> 00:03:02,420
它只负责一件事情

81
00:03:02,420 --> 00:03:05,260
就是把听到的声音转成文字

82
00:03:05,260 --> 00:03:06,940
嗯就单纯听写而已

83
00:03:06,940 --> 00:03:07,400
没错

84
00:03:07,400 --> 00:03:11,180
接着这个文字和音讯会被丢给Diarization

85
00:03:11,180 --> 00:03:13,680
也就是说话者分离系统

86
00:03:13,680 --> 00:03:15,200
等等说话者分离

87
00:03:15,200 --> 00:03:18,420
这听起来就像是一个事后侦探也

88
00:03:18,420 --> 00:03:19,200
对对对

89
00:03:19,200 --> 00:03:21,160
它就是负责回头去分辨说

90
00:03:21,160 --> 00:03:24,680
现在这句话到底是会议室里面的哪个人说的

91
00:03:24,680 --> 00:03:27,300
然后最后还需要过一个叫做

92
00:03:27,300 --> 00:03:29,700
endpointing的端点侦测系统

93
00:03:29,700 --> 00:03:30,760
端点侦测

94
00:03:30,760 --> 00:03:32,180
这又是做什么的

95
00:03:32,180 --> 00:03:33,860
这个就是用来判断说

96
00:03:33,860 --> 00:03:36,080
这个人到底讲完话了没有

97
00:03:36,080 --> 00:03:38,280
系统该不该切断麦克风

98
00:03:38,280 --> 00:03:40,420
哇这听起来就像是一条

99
00:03:40,420 --> 00:03:43,540
非常老旧而且缺乏效率的工厂管线

100
00:03:43,540 --> 00:03:45,960
真的非常贴切的比喻

101
00:03:45,960 --> 00:03:49,000
你可以想象就是三个工人站在输送袋旁边

102
00:03:49,000 --> 00:03:50,720
第一个工人只顾着

103
00:03:50,720 --> 00:03:52,440
听声音写字

104
00:03:52,440 --> 00:03:53,920
然后把写好的纸条

105
00:03:53,920 --> 00:03:55,420
赶快传给第二个工人

106
00:03:55,420 --> 00:03:56,560
然后第二个工人

107
00:03:56,560 --> 00:03:57,840
再去对比声音

108
00:03:57,840 --> 00:03:59,260
猜这到底是谁说的

109
00:03:59,260 --> 00:04:01,280
最后再传给第三个工人

110
00:04:01,280 --> 00:04:02,360
去决定这句话

111
00:04:02,360 --> 00:04:03,580
是不是真的结束了

112
00:04:03,580 --> 00:04:04,060
没错

113
00:04:04,060 --> 00:04:05,300
就是这么繁琐

114
00:04:05,300 --> 00:04:07,440
而且只要其中一个环节出错

115
00:04:07,440 --> 00:04:09,720
整个流程就会大塞车

116
00:04:09,720 --> 00:04:10,140
真的

117
00:04:10,140 --> 00:04:11,280
我之前用过一些

118
00:04:11,280 --> 00:04:12,400
会议记录软体

119
00:04:12,400 --> 00:04:14,240
最常发生的惨剧就是

120
00:04:14,240 --> 00:04:15,000
那个

121
00:04:15,000 --> 00:04:17,680
A老板的话才刚讲到一半

122
00:04:17,680 --> 00:04:20,600
然后旁边B员工稍微咳嗽了一声

123
00:04:20,600 --> 00:04:22,580
哦对系统就切断了

124
00:04:22,580 --> 00:04:24,860
没错系统就以为A讲完了

125
00:04:24,860 --> 00:04:27,160
直接把句子无情切断

126
00:04:27,160 --> 00:04:29,240
然后还把后面半句算在B头上

127
00:04:29,240 --> 00:04:30,620
这超尴尬的

128
00:04:30,620 --> 00:04:32,740
这真的是传统架构的致命伤

129
00:04:32,740 --> 00:04:36,700
过去这种做法通常需要大量的事后处理

130
00:04:36,700 --> 00:04:38,700
也就是等你整段话都录完了

131
00:04:38,700 --> 00:04:41,640
系统才能开始跑这些复杂的对比流程

132
00:04:41,640 --> 00:04:44,600
所以根本不可能做到真正的及时

133
00:04:44,600 --> 00:04:48,780
所以Meta这次的做法是怎么解决这个问题的

134
00:04:48,780 --> 00:04:51,880
他们这次的做法可以说是破坏性的创新

135
00:04:51,880 --> 00:04:56,560
他们把三个工人的工作全部交给了一个超级大脑同时完成

136
00:04:56,560 --> 00:04:58,440
一个大脑全包了

137
00:04:58,440 --> 00:04:58,660
对

138
00:04:58,660 --> 00:05:00,240
Muse Voice Transcribe

139
00:05:00,240 --> 00:05:07,080
它其实是建构在Meta的Muse Spark家族之上的一款自回归多模态模型

140
00:05:07,080 --> 00:05:08,600
等一下先暂停一下

141
00:05:08,600 --> 00:05:11,380
那个自回归多模态模型

142
00:05:11,380 --> 00:05:15,120
这几个字听起来真的很像某种外星语言也

143
00:05:15,120 --> 00:05:18,160
如果我们要用最白话的方式来理解

144
00:05:18,160 --> 00:05:19,780
这到底是什么意思

145
00:05:19,780 --> 00:05:21,140
简单来说了

146
00:05:21,140 --> 00:05:22,580
多摩泰意味着

147
00:05:22,580 --> 00:05:25,380
这个模型它可以同时看懂文字

148
00:05:25,380 --> 00:05:26,920
也能听懂声音

149
00:05:26,920 --> 00:05:29,660
它不是只吃单一一种资料

150
00:05:29,660 --> 00:05:33,300
哦 就是眼观四面耳听八方的感觉

151
00:05:33,300 --> 00:05:33,720
没错

152
00:05:33,720 --> 00:05:35,100
那自回归的话

153
00:05:35,100 --> 00:05:37,260
就像是我们在玩文字接龙一样

154
00:05:37,260 --> 00:05:39,720
它会根据已经接收到的资讯

155
00:05:39,720 --> 00:05:41,940
不断去预测下一步是什么

156
00:05:41,940 --> 00:05:44,500
但最关键的突破在于

157
00:05:44,500 --> 00:05:47,060
Meta在同一个音讯串流里面

158
00:05:47,060 --> 00:05:49,120
直接引入了特殊的标记

159
00:05:49,120 --> 00:05:50,420
也就是TOKENS

160
00:05:50,420 --> 00:05:52,820
这完全打破了我们刚才说的

161
00:05:52,820 --> 00:05:54,340
那个三阶段管线

162
00:05:54,340 --> 00:05:54,920
标记

163
00:05:54,920 --> 00:05:57,340
在在实际运作的画面上

164
00:05:57,340 --> 00:05:58,840
到底会长什么样子

165
00:05:58,840 --> 00:05:59,860
你想象一下

166
00:05:59,860 --> 00:06:01,560
有一条连续的时间轴

167
00:06:01,560 --> 00:06:03,760
当模型在及时聆听的时候

168
00:06:03,760 --> 00:06:05,860
只要它侦测到换人说话了

169
00:06:05,860 --> 00:06:08,020
它不会去呼叫别的系统来帮忙

170
00:06:08,020 --> 00:06:08,900
那它怎么做

171
00:06:08,900 --> 00:06:11,940
它会直接在自己生成的文字流里面

172
00:06:11,940 --> 00:06:13,600
插入一个特殊的标记

173
00:06:13,600 --> 00:06:15,280
叫做Start of Turn

174
00:06:15,280 --> 00:06:17,380
也就是换人说话的标记

175
00:06:17,380 --> 00:06:18,980
然后紧接着标注

176
00:06:18,980 --> 00:06:21,260
这句话是Speaker A说的

177
00:06:21,260 --> 00:06:23,860
哇 直接在同一个时间轴上标出来

178
00:06:23,860 --> 00:06:24,820
对对对

179
00:06:24,820 --> 00:06:27,580
同样的 刚你停顿或者讲完的时候

180
00:06:27,580 --> 00:06:31,660
它也会立刻放入一个Speech Endpoint的语音结束标记

181
00:06:31,660 --> 00:06:33,960
等于说所有的文字内容

182
00:06:33,960 --> 00:06:36,580
还有身份辨识 语气停顿

183
00:06:36,580 --> 00:06:38,660
全都在同一个时间轴上

184
00:06:38,660 --> 00:06:41,820
由这个单一模型一次性全部吐出来

185
00:06:41,820 --> 00:06:43,020
没有错就是这样

186
00:06:43,020 --> 00:06:48,420
哇所以这代表着我们再也不用忍受那种很尴尬的语音输入时刻了耶

187
00:06:48,420 --> 00:06:52,440
你们一定有碰过平时对着手机讲话稍微停顿想个词

188
00:06:52,440 --> 00:06:54,060
系统就以为你讲完了

189
00:06:54,060 --> 00:06:56,120
对然后就无情的把你切断

190
00:06:56,120 --> 00:07:00,840
没错现在这个模型它能真正理解对话的节奏

191
00:07:00,840 --> 00:07:06,200
它知道你只是在脑袋里面那个卡词了而不是真的结束发言

192
00:07:06,200 --> 00:07:10,320
没错,这正是单一串流模型带来的巨大优势

193
00:07:10,320 --> 00:07:15,780
但其实,这背后隐藏着一个极度矛盾的工程难题

194
00:07:15,780 --> 00:07:17,020
怎么说?

195
00:07:17,480 --> 00:07:19,320
既然这是一个超级模型

196
00:07:19,320 --> 00:07:24,140
要及时处理所有的声音、身份和文字

197
00:07:24,140 --> 00:07:27,980
那他怎么知道什么时候该把文字吐出来?

198
00:07:28,460 --> 00:07:29,340
哦,对耶

199
00:07:29,340 --> 00:07:32,100
如果他一听到声音就立刻输出文字

200
00:07:32,100 --> 00:07:34,280
那肯定很容易听错嘛

201
00:07:34,280 --> 00:07:36,680
但如果他为了追求完美

202
00:07:36,680 --> 00:07:39,220
等听完整句话才输出

203
00:07:39,220 --> 00:07:41,440
那又失去了我们刚才说的

204
00:07:41,440 --> 00:07:42,720
及时的意义

205
00:07:42,720 --> 00:07:43,760
这真的很矛盾

206
00:07:43,760 --> 00:07:46,280
这就像我们人类大脑

207
00:07:46,280 --> 00:07:47,840
在听别人说话一样

208
00:07:47,840 --> 00:07:48,920
你想想看

209
00:07:48,920 --> 00:07:50,720
如果有人对我说苹果

210
00:07:50,720 --> 00:07:52,820
就很简单的日常词汇

211
00:07:52,820 --> 00:07:55,240
我的大脑几乎不需要思考

212
00:07:55,240 --> 00:07:56,520
马上就能写下来

213
00:07:56,520 --> 00:07:57,380
对 瞄懂

214
00:07:57,380 --> 00:07:58,700
但如果今天

215
00:07:58,700 --> 00:08:00,000
对方正在讲一段话

216
00:08:00,000 --> 00:08:10,240
但如果今天,對方正在講一個超級複雜的醫學專有名詞,或者是他講話含糊不清,甚至口音很重,我的大腦其實會稍微,呃,猶豫個零點幾秒。

217
00:08:10,240 --> 00:08:11,600
沒錯,會卡一下。

218
00:08:11,600 --> 00:08:19,820
对,就是要等它把所有的音节都发音完整了,然后我在脑海中拼错确认之后,我才敢写下来。

219
00:08:20,220 --> 00:08:24,040
这个AI模型难道也会模仿人类的这种犹豫吗?

220
00:08:24,360 --> 00:08:26,740
这个人类大脑的类比非常到位耶。

221
00:08:27,240 --> 00:08:31,160
其实META的工程师就是循着这个逻辑去设计的。

222
00:08:31,680 --> 00:08:36,800
他们在技术摆笔书里面把这个称之为自适应延迟技术。

223
00:08:37,160 --> 00:08:40,340
自适应延迟,听起来蛮高级的。

224
00:08:40,340 --> 00:08:50,520
为了做到这点,模型在接收音讯的时候,它会把声音切成极小极小的区块,精确来说是80毫秒一个区块。

225
00:08:50,760 --> 00:08:53,020
80毫秒,这也太短了吧?

226
00:08:53,460 --> 00:08:59,980
超短的,也就是说在一秒钟之内,模型会接收到12.5个音讯区块。

227
00:09:00,660 --> 00:09:06,160
而在接收到每一个80毫秒区块的当下,模型都必须做出一个决定。

228
00:09:06,340 --> 00:09:07,020
什么决动?

229
00:09:07,020 --> 00:09:10,200
就是我是要现在立刻输出文字

230
00:09:10,200 --> 00:09:14,240
还是我要输出一个特殊的Next Audio标记

231
00:09:14,240 --> 00:09:15,600
告诉自己说

232
00:09:15,600 --> 00:09:18,340
我需要再多听下一个音讯巨块

233
00:09:18,340 --> 00:09:19,420
这很酷耶

234
00:09:19,420 --> 00:09:21,900
但他怎么知道自己该不该等啊

235
00:09:21,900 --> 00:09:26,240
工程师到底是怎么教懂一个AI学会去内心的呢

236
00:09:26,240 --> 00:09:29,240
这里的关键就在于强化学习

237
00:09:29,240 --> 00:09:31,180
工程师在训练它的时候

238
00:09:31,180 --> 00:09:34,240
设计了一个非常严苛的奖励机制

239
00:09:34,240 --> 00:09:35,280
有多严苛

240
00:09:35,280 --> 00:09:39,900
他们把那个字错率的惩罚跟延迟时间的惩罚

241
00:09:39,900 --> 00:09:41,860
用惩罚结合在一起

242
00:09:41,860 --> 00:09:44,480
你可以想象这是一场极限挑战

243
00:09:44,480 --> 00:09:46,280
如果模型当个快手

244
00:09:46,280 --> 00:09:47,800
一听到声音就乱拆

245
00:09:47,800 --> 00:09:49,260
导致疯狂写错字

246
00:09:49,260 --> 00:09:50,560
他就会被扣分

247
00:09:50,560 --> 00:09:53,140
那如果他想当个腕面主义者呢

248
00:09:53,140 --> 00:09:53,740
哇

249
00:09:53,740 --> 00:09:58,080
如果他为了一个字等了两三秒才输说文字

250
00:09:58,080 --> 00:10:00,280
他一样会被狠狠扣分

251
00:10:00,280 --> 00:10:01,120
哇

252
00:10:01,120 --> 00:10:05,360
这就像是在训练一个极度严格的听打员

253
00:10:05,360 --> 00:10:07,260
你就是不能错

254
00:10:07,260 --> 00:10:08,760
但是你也不能慢

255
00:10:08,760 --> 00:10:10,820
就是这种严格的训练

256
00:10:10,820 --> 00:10:13,320
逼迫模型学会了去判断语境

257
00:10:13,320 --> 00:10:15,520
碰到简单的日常对话

258
00:10:15,520 --> 00:10:17,040
它就立刻秒翻

259
00:10:17,040 --> 00:10:19,420
碰到尖色的专有名词

260
00:10:19,420 --> 00:10:21,200
或者是模糊的发音

261
00:10:21,200 --> 00:10:24,520
它就会选择多等个几十毫秒

262
00:10:24,520 --> 00:10:27,920
收集更多声音特征之后再输出

263
00:10:27,920 --> 00:10:28,960
这蛮合理的

264
00:10:28,960 --> 00:10:29,520
对

265
00:10:29,520 --> 00:10:31,120
在工程界

266
00:10:31,120 --> 00:10:33,480
这被称回达到了Pareto Front

267
00:10:33,480 --> 00:10:36,480
也就是在速度与准确度之间

268
00:10:36,480 --> 00:10:39,500
找到了无法再被超越的最佳边界

269
00:10:39,500 --> 00:10:41,360
最终的结果是

270
00:10:41,360 --> 00:10:42,840
Muse Voice Transcribe

271
00:10:42,840 --> 00:10:44,440
从你讲完话

272
00:10:44,440 --> 00:10:46,800
到萤幕上出现正确文字

273
00:10:46,800 --> 00:10:49,520
只需要惊人的0.16秒

274
00:10:49,520 --> 00:10:50,040
等等

275
00:10:50,040 --> 00:10:53,140
我要在这里稍微挑战一下

276
00:10:53,140 --> 00:10:54,960
0.16秒

277
00:10:54,960 --> 00:10:57,340
人类大脑神经传导

278
00:10:57,340 --> 00:10:58,720
做出反应的时间

279
00:10:58,720 --> 00:11:00,240
搞不好都比这个慢

280
00:11:00,240 --> 00:11:03,960
一个AI模型要在这0.16秒内

281
00:11:03,960 --> 00:11:05,100
切分音讯

282
00:11:05,100 --> 00:11:06,740
决定要不要等待

283
00:11:06,740 --> 00:11:09,040
还要辨识是谁说话

284
00:11:09,040 --> 00:11:10,380
并伸出文字

285
00:11:10,380 --> 00:11:13,920
他怎么可能在这么短的时间内运算完

286
00:11:13,920 --> 00:11:15,820
却不吃垮电脑的效能

287
00:11:15,820 --> 00:11:18,520
这真的是个非常好的质疑

288
00:11:18,520 --> 00:11:20,340
直觉上做这么多事情

289
00:11:20,340 --> 00:11:22,160
应该会严重拖慢系统

290
00:11:22,160 --> 00:11:23,060
对啊

291
00:11:23,060 --> 00:11:24,600
听起来运算量超大的

292
00:11:24,600 --> 00:11:26,660
但实际情况恰好相反

293
00:11:26,660 --> 00:11:29,060
正因为它舍去了过去那种

294
00:11:29,060 --> 00:11:32,580
三套系统互相传递资料的管线架构

295
00:11:32,580 --> 00:11:35,380
它省去了系统之间来回沟通

296
00:11:35,380 --> 00:11:38,300
和对时间轴的庞大运算成本

297
00:11:38,300 --> 00:11:40,800
就不用大家互相等来等去

298
00:11:40,800 --> 00:11:41,400
没错

299
00:11:41,400 --> 00:11:43,880
一个单一的自回归模型

300
00:11:43,880 --> 00:11:46,280
只要硬体的记忆体频宽足够

301
00:11:46,280 --> 00:11:48,560
它处理连续串流的效率

302
00:11:48,560 --> 00:11:51,740
其实远比传统的分工合作还要快得多

303
00:11:51,740 --> 00:11:53,240
这也就是为什么

304
00:11:53,240 --> 00:11:56,960
它能压榨出0.16秒这种极限速度

305
00:11:56,960 --> 00:11:58,680
哇 原来是这样

306
00:11:58,680 --> 00:12:02,280
理论和底层机制听起来确实很惊人啦

307
00:12:02,280 --> 00:12:03,640
但我们都知道

308
00:12:03,640 --> 00:12:06,720
实验室里面的数据是一回事

309
00:12:06,720 --> 00:12:09,620
把模型丢进残酷的真实世界

310
00:12:09,620 --> 00:12:12,120
面对各种混乱的对话场景

311
00:12:12,120 --> 00:12:14,000
那又是另外一回事了嘛

312
00:12:14,000 --> 00:12:16,960
它在真实环境中的表现到底怎么样

313
00:12:16,960 --> 00:12:20,040
那我们就来看看知名的AI评测机构

314
00:12:20,040 --> 00:12:23,200
Artificial Analysis给出的硬数据吧

315
00:12:23,200 --> 00:12:26,900
在纯英文的即时串流测试里面

316
00:12:26,900 --> 00:12:31,240
Muse Voice Transcribe的字错率仅仅只有3.1%

317
00:12:31,240 --> 00:12:32,480
3.1%

318
00:12:32,480 --> 00:12:34,460
这算是很好的成绩吗

319
00:12:34,460 --> 00:12:35,560
非常好

320
00:12:35,560 --> 00:12:36,700
这是什么概面

321
00:12:36,700 --> 00:12:40,140
它直接击败了目前语音领域的王者

322
00:12:40,140 --> 00:12:42,180
也就是那个11实验室

323
00:12:42,180 --> 00:12:44,400
他们的3.6%

324
00:12:44,400 --> 00:12:45,940
哦 11 labs

325
00:12:45,940 --> 00:12:46,480
对

326
00:12:46,480 --> 00:12:51,360
它也赢过了Oper AI的GPT Live Transcribe的3.9

327
00:12:51,360 --> 00:12:54,260
甚至把Google在8月才刚发布的

328
00:12:54,260 --> 00:12:56,900
German Live 3.5 Transcribe Live的4.0

329
00:12:56,900 --> 00:12:58,420
都抛在脑后

330
00:12:58,420 --> 00:13:00,820
哇 全部斩压耶

331
00:13:00,820 --> 00:13:01,340
没错

332
00:13:01,340 --> 00:13:04,300
而且在讲者辨识错误率上

333
00:13:04,300 --> 00:13:06,500
它的错误率只有17.5%

334
00:13:06,500 --> 00:13:09,320
这个数字甚至赢过了市面上许多

335
00:13:09,320 --> 00:13:13,080
允许你事后慢慢处理的那种非即时模型

336
00:13:13,080 --> 00:13:15,540
不过我觉得我们必须拉回现实一下

337
00:13:15,540 --> 00:13:17,640
我们的听众最关心的

338
00:13:17,640 --> 00:13:20,700
绝对不是在一个纯英文

339
00:13:20,700 --> 00:13:23,160
大家都自证强缘的环境里面

340
00:13:23,160 --> 00:13:23,900
它有多准

341
00:13:23,900 --> 00:13:24,980
嗯 确实

342
00:13:24,980 --> 00:13:25,480
对嘛

343
00:13:25,480 --> 00:13:26,520
我们在排湾

344
00:13:26,520 --> 00:13:29,420
平时讲话可是中阴夹杂的耶

345
00:13:29,420 --> 00:13:30,540
像是那个

346
00:13:30,540 --> 00:13:31,800
我明天有个meeting

347
00:13:31,800 --> 00:13:33,460
记得remind我一下

348
00:13:33,460 --> 00:13:36,340
这种句子才是我们最正式的日常啊

349
00:13:36,340 --> 00:13:38,940
Meta其实在这方面下了很大的功夫

350
00:13:38,940 --> 00:13:40,840
这款模型发布的时候

351
00:13:40,840 --> 00:13:43,160
就已经支援了70多种语言

352
00:13:43,160 --> 00:13:46,200
其中有25种是经过完整验证的

353
00:13:46,200 --> 00:13:47,380
最厉害的是

354
00:13:47,380 --> 00:13:51,280
它原生就支援句子内部的鱼马转换

355
00:13:51,280 --> 00:13:52,140
鱼马转换

356
00:13:52,140 --> 00:13:55,220
意思是它不需要我去切换那个语言设定

357
00:13:55,220 --> 00:13:56,780
就能自动听懂吗

358
00:13:56,780 --> 00:13:57,100
是的

359
00:13:57,100 --> 00:13:59,740
在官方释出的测试来源里面

360
00:13:59,740 --> 00:14:02,820
有一段充满硬核科技术语的中文录音

361
00:14:02,820 --> 00:14:04,080
测试者说

362
00:14:04,080 --> 00:14:06,980
用Olama跑了向Meta的Muse Glamour

363
00:14:06,980 --> 00:14:11,540
直接deploy在我的NVIDIA RTX 3090上

364
00:14:11,540 --> 00:14:14,960
我这句话里面充满了英文专有名词数字

365
00:14:14,960 --> 00:14:18,320
而且中英文切换完全是随机的耶

366
00:14:18,320 --> 00:14:18,720
没错

367
00:14:18,720 --> 00:14:21,860
母亲不需要任何提示就能无缝辨识

368
00:14:21,860 --> 00:14:24,020
还能根据文法的上下文

369
00:14:24,020 --> 00:14:25,780
去自动校正评字

370
00:14:25,780 --> 00:14:27,580
说到真实世界的混乱

371
00:14:27,580 --> 00:14:29,080
我看到资料里面

372
00:14:29,080 --> 00:14:31,800
还有一个让我下巴掉下来的测试

373
00:14:31,800 --> 00:14:33,200
你说那个爬山的那个吗

374
00:14:33,200 --> 00:14:33,840
对对对

375
00:14:33,840 --> 00:14:36,260
那个是一个长达一小时的音档

376
00:14:36,260 --> 00:14:39,140
里面竟然有11个人

377
00:14:39,140 --> 00:14:41,040
在洛杉矶讨论爬山

378
00:14:41,040 --> 00:14:43,500
甚至还聊到遇到土狼的经历

379
00:14:43,500 --> 00:14:44,660
你想想看

380
00:14:44,660 --> 00:14:47,080
11个人在闲聊耶

381
00:14:47,080 --> 00:14:48,660
一定有人在大笑

382
00:14:48,660 --> 00:14:50,040
有人互相抢话

383
00:14:50,040 --> 00:14:52,340
背景绝对还有各种环境噪音

384
00:14:52,340 --> 00:14:53,420
这对AI来说

385
00:14:53,420 --> 00:14:55,040
根本是地狱级的考验

386
00:14:55,040 --> 00:14:55,580
真的

387
00:14:55,580 --> 00:14:58,220
结果系统竟然能精准的

388
00:14:58,220 --> 00:15:00,040
把那11个人的发言

389
00:15:00,040 --> 00:15:02,120
一次不漏的拆解开来

390
00:15:02,120 --> 00:15:04,280
还标上speakerA到K

391
00:15:04,280 --> 00:15:06,900
这如果叫人类的听答员来做

392
00:15:06,900 --> 00:15:08,140
听个10分钟

393
00:15:08,140 --> 00:15:09,740
绝对崩溃磁石吧

394
00:15:09,740 --> 00:15:11,360
但这在技术上

395
00:15:11,360 --> 00:15:12,740
到底是怎么办到的

396
00:15:12,740 --> 00:15:13,820
如果两个人

397
00:15:13,820 --> 00:15:16,060
在同一微秒重叠讲话呢

398
00:15:16,060 --> 00:15:17,380
这就是我们刚刚说的

399
00:15:17,380 --> 00:15:20,180
多模态标记发挥威力的地方了

400
00:15:20,180 --> 00:15:22,520
当两个人同时讲话的时候

401
00:15:22,520 --> 00:15:25,560
传统系统的音轨会整个混在一起

402
00:15:25,560 --> 00:15:27,560
变成无法辨识的噪音

403
00:15:27,560 --> 00:15:29,680
嗯就胡称一团

404
00:15:29,680 --> 00:15:30,140
对

405
00:15:30,140 --> 00:15:31,940
但Muse Voice Transcribe

406
00:15:31,940 --> 00:15:35,440
因为是从极短的80毫秒音讯区块

407
00:15:35,440 --> 00:15:36,460
去提取特征

408
00:15:36,460 --> 00:15:39,440
它可以同时触发多个讲者的token

409
00:15:39,440 --> 00:15:41,060
可以同时触发

410
00:15:41,060 --> 00:15:41,600
没错

411
00:15:41,600 --> 00:15:44,460
它会在文字流里面并行输出

412
00:15:44,460 --> 00:15:46,300
Speaker A说了这一段

413
00:15:46,300 --> 00:15:48,660
同时Speaker B正在说那一段

414
00:15:48,660 --> 00:15:50,500
它不是单线程的处理

415
00:15:50,500 --> 00:15:52,460
而是能像解开毛线球一样

416
00:15:52,460 --> 00:15:55,020
把交织在一起的音讯分离出来

417
00:15:55,020 --> 00:15:56,920
这技术真的蛮扯的

418
00:15:56,920 --> 00:15:59,760
不过身为替台湾听众把关的酒色

419
00:16:04,500 --> 00:16:06,000
哦,什麼事實?

420
00:16:06,000 --> 00:16:14,700
就是,根據台灣知名科技媒體電腦王阿達的實際測試,這項技術目前有一個在地化的致命缺陷。

421
00:16:14,700 --> 00:16:16,100
嗯,在地化

422
00:16:16,100 --> 00:16:20,660
对,尽管你对着麦克风说的是很标准的台湾繁体中文

423
00:16:20,660 --> 00:16:26,040
它目前在荧幕上输出的文字依然全都是,呃,简体中文

424
00:16:26,040 --> 00:16:28,140
这确实蛮麻烦的

425
00:16:28,140 --> 00:16:32,560
对啊,不管设定怎么调,目前就是无法原声输出繁体字

426
00:16:32,560 --> 00:16:36,600
这对台湾的使用者来说绝对还是一个不小的痛点啊

427
00:16:36,600 --> 00:16:39,440
这的确是一个很明显的在地化缺陷

428
00:16:39,440 --> 00:16:43,140
其实这显示出虽然模型的听力没有问题

429
00:16:43,140 --> 00:16:46,740
它的神经网络完全能理解中文的语义

430
00:16:46,740 --> 00:16:48,260
听得懂但写出来不对

431
00:16:48,260 --> 00:16:48,720
没错

432
00:16:48,720 --> 00:16:51,780
但在最后的文字渲染与输出层面

433
00:16:51,780 --> 00:16:53,360
META的训练资料库

434
00:16:53,360 --> 00:16:56,160
可能还是过度向简体中文倾斜了

435
00:16:56,160 --> 00:16:58,880
这也反映了全球化大语言模型

436
00:16:58,880 --> 00:16:59,860
在落地的时候

437
00:16:59,860 --> 00:17:03,620
必然会面临到的文化与区域细节挑战了

438
00:17:03,620 --> 00:17:05,400
这部分就需要META后续

439
00:17:05,400 --> 00:17:06,820
透过微调来解决了

440
00:17:06,820 --> 00:17:08,640
既然提到META的策略

441
00:17:08,640 --> 00:17:11,460
我们来聊聊一个很反常的商业布局好了

442
00:17:11,460 --> 00:17:12,540
我们都知道

443
00:17:12,540 --> 00:17:16,320
Meta过去这一年来一直高举着开源的大旗

444
00:17:16,320 --> 00:17:19,920
把他们最大的AI模型免费开放给全世界

445
00:17:19,920 --> 00:17:21,600
对 开源界的老大哥

446
00:17:21,600 --> 00:17:22,300
没错

447
00:17:22,300 --> 00:17:26,620
但这一次Mail Voice Transcribe却选择了封闭权重

448
00:17:26,620 --> 00:17:28,800
也就是不开源了

449
00:17:28,800 --> 00:17:29,380
为什么

450
00:17:29,380 --> 00:17:32,220
这次Meta确实选择了一条不一样的路

451
00:17:32,220 --> 00:17:33,600
如果你是开发者

452
00:17:33,600 --> 00:17:35,780
你无法直接下载模型版身

453
00:17:35,780 --> 00:17:38,760
你必须透过Meta Model API来使用

454
00:17:38,760 --> 00:17:39,980
哦 要收费了

455
00:17:39,980 --> 00:17:42,080
不过收费其实相当低廉啦

456
00:17:42,080 --> 00:17:45,860
大约是每1000分钟的音讯收费3美元

457
00:17:45,860 --> 00:17:49,720
换算下来一个小时大概才0.18美元

458
00:17:49,720 --> 00:17:51,520
那真的蛮便宜的

459
00:17:51,520 --> 00:17:52,060
对

460
00:17:52,060 --> 00:17:56,680
另外一般使用者其实可以透过他们新推出的程式撰写工具

461
00:17:56,680 --> 00:17:58,020
叫做Muse Code

462
00:17:58,020 --> 00:18:03,500
或者是直接在Mac电脑上下载Meta AI的应用程式来使用这项技术

463
00:18:03,500 --> 00:18:05,700
所以这到底意味着什么

464
00:18:05,700 --> 00:18:11,620
Meta把这个听力超强的模型紧紧锁在自家的生态系里面

465
00:18:11,620 --> 00:18:14,600
甚至还跟Mac的键盘深度绑定

466
00:18:14,600 --> 00:18:18,580
Mark Zuckerberg到底在盘算什么更大的棋局

467
00:18:18,580 --> 00:18:22,520
这不仅仅是为了赚那个0.18美元吧

468
00:18:22,520 --> 00:18:25,180
绝对不是为了这点API的零头啦

469
00:18:25,180 --> 00:18:28,460
如果我们深入观察 Zuckerberg近期的内部发言

470
00:18:28,460 --> 00:18:32,300
他不断强调未来的AI必须是个人化的超级智慧

471
00:18:32,300 --> 00:18:33,980
个人化的超级智慧

472
00:18:33,980 --> 00:18:36,120
这听起来很抽象耶

473
00:18:36,120 --> 00:18:40,100
这跟一个语音转文字的模型有什么直接关系

474
00:18:40,100 --> 00:18:41,340
关系太大了

475
00:18:41,340 --> 00:18:43,920
如果AI要成为你最贴身

476
00:18:43,920 --> 00:18:45,420
最懂你的个人助理

477
00:18:45,420 --> 00:18:46,720
它就不能只是一个

478
00:18:46,720 --> 00:18:47,900
你对着手机大喊

479
00:18:47,900 --> 00:18:49,220
嘿帮我设定闹钟

480
00:18:49,220 --> 00:18:50,900
那种死板的声控工具

481
00:18:50,900 --> 00:18:53,580
Meta的最终目标是硬体

482
00:18:53,580 --> 00:18:54,840
特别是他们正在

483
00:18:54,840 --> 00:18:56,740
大力发展的AI视觉眼镜

484
00:18:56,740 --> 00:18:58,200
哦 AI眼镜

485
00:18:58,200 --> 00:18:58,960
你试想一下

486
00:18:58,960 --> 00:19:00,660
你戴着AI眼镜的一天好了

487
00:19:00,660 --> 00:19:02,840
你走在台北吵杂的街头

488
00:19:02,840 --> 00:19:04,260
旁边有车流声

489
00:19:04,260 --> 00:19:06,260
你正在点一杯珍珠奶茶

490
00:19:06,260 --> 00:19:08,160
店员问你要几分甜

491
00:19:08,160 --> 00:19:09,700
这时你旁边的朋友

492
00:19:09,700 --> 00:19:10,780
突然插花问你

493
00:19:10,780 --> 00:19:11,940
等下要去哪里

494
00:19:11,940 --> 00:19:14,200
哇 这个场景非常真实

495
00:19:14,200 --> 00:19:15,620
也非常混乱

496
00:19:15,620 --> 00:19:17,580
对 如果你的AI眼镜

497
00:19:17,580 --> 00:19:19,000
没有完美的机器听觉

498
00:19:19,000 --> 00:19:20,100
它就只是一个

499
00:19:20,100 --> 00:19:21,920
戴在脸上的笨重摄影机

500
00:19:21,920 --> 00:19:23,960
但有了Muse Voice Transcribe

501
00:19:23,960 --> 00:19:26,020
眼睛能在0.16秒内

502
00:19:26,020 --> 00:19:27,820
分辨出哪句是电源说的

503
00:19:27,820 --> 00:19:29,420
哪句是朋友说的

504
00:19:29,420 --> 00:19:31,160
还能过滤掉车流声

505
00:19:31,160 --> 00:19:33,120
没错 甚至刚你转头

506
00:19:33,120 --> 00:19:34,920
忘记朋友刚才问什么的时候

507
00:19:34,920 --> 00:19:36,740
AI已经把答案准备好

508
00:19:36,740 --> 00:19:38,320
投射在你的镜片上了

509
00:19:38,320 --> 00:19:40,040
哇 这画面感很强耶

510
00:19:40,040 --> 00:19:44,220
这就是为什么Meta现在让你只要在Mac电脑上按住FN键

511
00:19:44,220 --> 00:19:47,800
就能在任何视窗进行痊愈的及时听写

512
00:19:47,800 --> 00:19:50,200
这不是一个单纯的打字工具

513
00:19:50,200 --> 00:19:53,780
这是Meta在训练你习惯与AI无缝对话

514
00:19:53,780 --> 00:19:58,320
为未来的穿戴式硬体生态系铺设最底层的基础设施

515
00:19:58,320 --> 00:20:02,280
今天的探讨真的把很多点都串起来了

516
00:20:02,280 --> 00:20:05,800
我们现在正在见证一个历史性的转捩点

517
00:20:05,800 --> 00:20:09,880
机器在复杂环境下的听觉处理能力

518
00:20:09,880 --> 00:20:13,320
正在以前所未有的速度超越人类

519
00:20:13,320 --> 00:20:14,620
对,真的很快

520
00:20:14,620 --> 00:20:18,780
他能毫不犹豫地在0.16秒内处理资讯

521
00:20:18,780 --> 00:20:22,120
他能同时分辨十几个人的声音

522
00:20:22,120 --> 00:20:24,480
他甚至学会了像人类一样

523
00:20:24,480 --> 00:20:27,420
在困难的词汇上稍微犹豫一下

524
00:20:27,420 --> 00:20:31,440
没错,从过去笨重的三步骤拼装车系统

525
00:20:31,440 --> 00:20:34,560
进化到现在单一串流的超级模型

526
00:20:34,560 --> 00:20:37,840
这不仅仅是技术指标上的胜利而已

527
00:20:37,840 --> 00:20:41,720
更是人机互动方式的一次彻底革命

528
00:20:41,720 --> 00:20:42,900
真的是革命

529
00:20:42,900 --> 00:20:46,620
虽然目前还像是繁体中文资源不足这样的缺点

530
00:20:46,620 --> 00:20:49,720
但这项技术为未来AI眼镜

531
00:20:49,720 --> 00:20:52,600
和无缝语音互动所打下的基础

532
00:20:52,600 --> 00:20:55,900
其实已经预示了科技的下一步走向了

533
00:20:55,900 --> 00:20:56,220
真的

534
00:20:56,220 --> 00:20:59,280
在我们结束今天的深度探讨之前

535
00:20:59,280 --> 00:21:02,560
我想留给你们一个发人深审的问题

536
00:21:02,560 --> 00:21:04,080
想象一下

537
00:21:04,080 --> 00:21:06,320
如果未来的每一台装置

538
00:21:06,320 --> 00:21:07,700
每一副眼镜

539
00:21:07,700 --> 00:21:09,800
都能在0.16秒内

540
00:21:09,800 --> 00:21:11,300
完美记录下

541
00:21:11,300 --> 00:21:13,160
20个人在会议室里

542
00:21:13,160 --> 00:21:14,160
互相强化

543
00:21:14,160 --> 00:21:16,760
甚至私下碎碎念的内容

544
00:21:16,760 --> 00:21:18,460
并精准标记

545
00:21:18,460 --> 00:21:19,740
到底是谁说的

546
00:21:19,740 --> 00:21:21,280
这对我们的社会

547
00:21:21,280 --> 00:21:22,520
到底意味着什么

548
00:21:22,520 --> 00:21:23,060
嗯

549
00:21:23,060 --> 00:21:23,960
这蛮可怕的

550
00:21:23,960 --> 00:21:24,620
对吧

551
00:21:24,620 --> 00:21:26,260
当过去那种

552
00:21:26,260 --> 00:21:28,860
听过就算了的随性闲聊

553
00:21:28,860 --> 00:21:31,380
全部变成可以被精准搜寻

554
00:21:31,380 --> 00:21:32,780
被永久记录的

555
00:21:32,780 --> 00:21:35,100
完美数位逐字稿的时候

556
00:21:35,100 --> 00:21:37,500
这会让我们的工作与生活

557
00:21:37,500 --> 00:21:38,840
变得极度高效

558
00:21:38,840 --> 00:21:41,340
还是会让我们在每次开口之前

559
00:21:41,340 --> 00:21:43,820
都变得极度焦虑多疑呢

560
00:21:43,820 --> 00:21:46,240
这是一个值得你在关掉节目后

561
00:21:46,240 --> 00:21:47,860
好好思考的问题

562
00:21:47,860 --> 00:21:49,160
谢谢收听

563
00:21:49,160 --> 00:21:50,360
欢迎订阅分享

564
00:21:50,360 --> 00:21:51,600
听书Podcast频道

565
00:21:51,600 --> 00:21:52,620
我们下次见