首页 > AI前沿 > Decode Early, Verify Late: Streaming Offline Speech Recognizers Without Accuracy Loss

Decode Early, Verify Late: Streaming Offline Speech Recognizers Without Accuracy Loss

arXiv机器学习 2026-11-15 23:15 12 阅读 查看原文
用户现在需要把给定的新闻正文排版成语义化的HTML片段,按照要求,首先看内容,都是自然段落?不对,首先第一段是介绍DEVL的,然后后面的部分,哦,先理清楚结构:首先第一句到offline ASR models是引入,然后后面讲框架原理,然后实验结果,然后性能表现。不对,要按照要求,每个自然段落用p,合适的用h2/h3划分小节,重要内容加strong,保留所有信息。 等下,首先,先看原文的结构: 第一部分:Voice agents...offline ASR models. 这是引出DEVL的,可以作为第一段?或者小节?不对,首先,开头的部分是背景和提出的框架,然后“The server keeps partial transcripts...exact arithmetic.”是讲框架的工作原理,然后“We adapt DEVL to four model families.”是适配情况,然后“In our five-recognizer comparison...official streaming mode.”是实验对比结果,然后“It is the only system...64 on WildASR.”是性能达标情况。 不对,要准确,不能增删内容。首先,先划分小节?比如: 首先,

框架提出

?不对,不能自己加内容?哦不对,h2是划分小节,但是小节标题的文本?不对哦,规则是保留原文所有信息,不得增删改内容。哦对,不能加原文没有的标题文字!哦对,那小节划分的话,是不是如果原文有明显的小节内容,用h2/h3,但是标题不能用自己加的,不对?哦不对,等下看规则2:“长文用

/

划分小节”,哦,但是如果原文没有小标题的话,是不是要合理划分,但是不能新增内容,所以小节的内容就是原文的段落,h2的文本?不对,不对,不能加原文没有的文字,所以h2是不是可以不用?不对,再看原文的内容,都是连贯的?不对,先仔细看: 原文的完整内容: 1. Voice agents need accurate transcripts soon after a user stops speaking. We propose Decode Early, Verify Late (DEVL), a training-free framework for streaming inference with offline ASR models. 2. The server keeps partial transcripts as drafts and verifies the latest against the complete audio in one forward pass at turn end. Under stated input and decoding conditions, the final equals the model's greedy offline output in exact arithmetic. 3. We adapt DEVL to four model families. 4. In our five-recognizer comparison, 95-100% of final transcripts are byte-identical to the offline transcripts, WER stays within 0.13 points of offline, and single-stream time to final falls 1.5-7.3x. 5. Serving Qwen3-ASR-1.7B on one A100, DEVL returns finals a median 22-26 ms after turn end at up to 64 streams, 2.7-3.9x sooner than the official streaming mode. 6. It is the only system we tested that meets a 100 ms p99 target, up to 80 streams on Pipecat-STT and 64 on WildASR. 不对,这些是自然段落吗?原文是一整段?不对,用户给的正文是一整段?看用户给的正文: 正文如下: Voice agents need accurate transcripts soon after a user stops speaking. We propose Decode Early, Verify Late (DEVL), a training-free framework for streaming inference with offline ASR models. The server keeps partial transcripts as drafts and verifies the latest against the complete audio in one forward pass at turn end. Under stated input and decoding conditions, the final equals the model's greedy offline output in exact arithmetic. We adapt DEVL to four model families. In our five-recognizer comparison, 95-100% of final transcripts are byte-identical to the offline transcripts, WER stays within 0.13 points of offline, and single-stream time to final falls 1.5-7.3x. Serving Qwen3-ASR-1.7B on one A100, DEVL returns finals a median 22-26 ms after turn end at up to 64 streams, 2.7-3.9x sooner than the official streaming mode. It is the only system we tested that meets a 100 ms p99 target, up to 80 streams on Pipecat-STT and 64 on WildASR. 哦,原来用户给的正文是一整段的纯文本,那需要划分成自然段落,每个用p,重要内容加strong,长文用h2/h3?不对,但是不能加原文没有的标题内容啊,哦,等下,规则2说“长文用

/

划分小节”,但是如果没有小标题的话,是不是可以将不同的主题内容用h2包裹?不对,h2是标题,不能自己造标题,那是不是可以不用h2/h3?不对,再仔细看规则,规则是“每个自然段落用

包裹,长文用

/

划分小节”,如果原文没有小节标题的话,那是不是合理划分自然段落,重要内容加strong。 首先,划分自然段落: 第一段(背景与框架提出):

Voice agents need accurate transcripts soon after a user stops speaking. We propose Decode Early, Verify Late (DEVL), a training-free framework for streaming inference with offline ASR models.

第二段(框架工作原理):

The server keeps partial transcripts as drafts and verifies the latest against the complete audio in one forward pass at turn end. Under stated input and decoding conditions, the final equals the model's greedy offline output in exact arithmetic.

第三段(适配情况):

We adapt DEVL to four model families.

第四段(多模型对比结果):

In our five-recognizer comparison, 95-100% of final transcripts are byte-identical to the offline transcripts, WER stays within 0.13 points of offline, and single-stream time to final falls 1.5-7.3x.

第五段(单卡部署性能):

Serving Qwen3-ASR-1.7B on one A100, DEVL returns finals a median 22-26 ms after turn end at up to 64 streams, 2.7-3.9x sooner than the official streaming mode.

第六段(性能达标情况):

It is the only system we tested that meets a 100 ms p99 target, up to 80 streams on Pipecat-STT and 64 on WildASR.

不对,但是要不要用h2?比如,框架原理、实验结果这些,但是不能用自己加的标题,那是不是不需要h2?因为规则说长文用h2/h3划分,但是如果原文没有小标题,就不用强行加,不然就增改内容了。对,因为不能增删改内容,所以h2/h3只有在原文有对应的小标题的时候才用,这里原文没有,所以只用p,重要内容加strong。 等下,再检查有没有增删内容:所有原文的内容都要保留,对吧?刚才的划分是把原文的内容按语义分成不同的自然段落,没有加任何内容,重要的地方加strong,比如DEVL的名称,关键数据,这些是合理的。 再调整下,有没有更合适的?比如,第四段的WER和时间数据要不要加strong?也可以,比如:

In our five-recognizer comparison, 95-100% of final transcripts are byte-identical to the offline transcripts, WER stays within 0.13 points of offline, and single-stream time to final falls 1.5-7.3x.

第五段的也可以把关键部分加strong,没错。 再检查所有内容都在: Voice agents need accurate transcripts soon after a user stops speaking. We propose Decode Early, Verify Late (DEVL), a training-free framework for streaming inference with offline ASR models. 都有。 The server keeps partial transcripts as drafts and verifies the latest against the complete audio in one forward pass at turn end. Under stated input and decoding conditions, the final equals the model's greedy offline output in exact arithmetic. 都有。 We adapt DEVL to four model families. 有。 In our five-recognizer comparison, 95-100% of final transcripts are byte-identical to