面向实时语音转写场景的流式与非流式语音识别性能对比研究

doi:10.12267/j.issn.2096-5931.2026.09.010

信息通信技术与政策 ›› 2026, Vol. 52 ›› Issue (9): 69-77.doi: 10.12267/j.issn.2096-5931.2026.09.010

面向实时语音转写场景的流式与非流式语音识别性能对比研究

Performance comparison of streaming and non-streaming speech recognition for real-time transcription scenarios

李鲲程¹, 葛宏伟¹, 郝航¹, 范春梅²

¹ 中国信息通信研究院信息管理中心, 北京 100191
² 北京邮电大学信息与通信工程学院, 北京 100876

收稿日期:2026-04-03 出版日期:2026-09-25 发布日期:2026-09-30
通讯作者: 范春梅
作者简介:
李鲲程,中国信息通信研究院信息管理中心高级工程师,主要从事人工智能系统架构设计与优化、前沿技术落地验证、信息化建设项目管理和数据分析等方面的工作；
葛宏伟,中国信息通信研究院信息管理中心项目主管,主要从事人工智能应用需求场景分析、应用系统设计与全流程管理等方面的工作；
郝航,中国信息通信研究院信息管理中心工程师,主要从事计算平台的规划建设与运维等方面的工作

LI Kuncheng¹, GE Hongwei¹, HAO Hang¹, FAN Chunmei²

¹ Information Management Center, China Academy of Information and Communications Technology, Beijing 100191, China
² School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, China

Received:2026-04-03 Online:2026-09-25 Published:2026-09-30
Contact: FAN Chunmei

摘要/Abstract

摘要：

为满足会议机器同传显示场景的需求,实时语音转写需满足低时延、高准确率、强可读性的要求。为此,通过构建系统化的评估指标体系和离线流式模拟评测方法,对比流式(Sherpa-onnx模型)与非流式(Faster-Whisper模型)两种语音识别方案的性能差异,为相关领域研究提供参考。试验结果表明,流式语音识别方案配合轻量级语义后处理策略,在整体性能表现上更适应工业级会议系统对实时语音转写的综合要求。

关键词: 实时语音转写, 滑动窗口, 语义后处理

Abstract:

To meet the requirements of real-time speech transcription display in conference simultaneous interpretation scenarios, the system must achieve low latency, high accuracy, and strong readability. For this purpose, systematic evaluation metrics and an offline streaming simulation assessment method are established to compare the performance differences between streaming (Sherpa-onnx) and non-streaming (Faster-Whisper) speech recognition solutions, offering a reference for research in related fields. Experimental results indicate that the streaming speech recognition approach, when coupled with a lightweight semantic post-processing strategy, exhibits overall performance that better satisfies the comprehensive demands of real-time speech transcription in industrial-grade conference systems.

Key words: real-time speech transcription, sliding window, semantic post-processing

中图分类号:

TP391
TN912

李鲲程, 葛宏伟, 郝航, 范春梅. 面向实时语音转写场景的流式与非流式语音识别性能对比研究[J]. 信息通信技术与政策, 2026, 52(9): 69-77.

LI Kuncheng, GE Hongwei, HAO Hang, FAN Chunmei. Performance comparison of streaming and non-streaming speech recognition for real-time transcription scenarios[J]. Information and Communications Technology and Policy, 2026, 52(9): 69-77.

导出引用管理器 EndNote|Ris|BibTeX

链接本文:

http://ictp.caict.ac.cn/CN/10.12267/j.issn.2096-5931.2026.09.010

http://ictp.caict.ac.cn/CN/Y2026/V52/I9/69

图/表 5

参考文献 19

[1]	Radford A, Kim J W, Xu T, et al. Robust speech recognition via large-scale weak supervision[PP/OL]. V1. arXiv (2022-12-06) [2026-04-03]. https://doi.org/10.48550/arXiv.2212.04356.
[2]	Macháček D, Dabre R, Bojar O. Turning whisper into real-time transcription system[PP/OL]. V1. arXiv (2023-07-27) [2026-04-03]. https://doi.org/10.48550/arXiv.2307.14743.
[3]	Wang H Y, Hu G Q, Lin G D, et al. Simul-whisper: attention-guided streaming whisper with truncation detection[PP/OL]. V1. arXiv ( 2024-06-24) [2026-04-03]. https://doi.org/10.21437/Interspeech.2024-1814.
[4]	Tsunoo E, Kashiwagi Y, Watanabe S. Streaming transformer asr with blockwise synchronous beam search[C]// 2021 IEEE Spoken Language Technology Workshop (SLT). Shenzhen: IEEE Press, 2021: 22-29.
[5]	Narayanan A, Prabhavalkar R, Chiu C C, et al. Recognizing long-form speech using streaming end-to-end models[C]// 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). Singapore: IEEE Press, 2019: 920-927.
[6]	Wu C Y, Wang Y Q, Shi Y Y, et al. Streaming transformer-based acoustic models using self-attention with augmented memory[PP/OL]. V1. arXiv (2020-05-16) [2026-04-03]. https://doi.org/10.48550/arXiv.2005.08042.
[7]	Yao Z W, Guo L Y, Yang X Y, et al. Zipformer: a faster and better encoder for automatic speech recognition[PP/OL]. V1. arXiv (2023-10-17) [2026-04-03]. https://doi.org/10.48550/arXiv.2310.11230.
[8]	Chen G G, Chai S Z, Wang G B, et al. Gigaspeech:an evolving, multi-domain asr corpus with 10,000 hours of transcribed audio[PP/OL]. V1. arXiv (2021-06-13) [2026-04-03]. https://doi.org/10.21437/Interspeech.2021-1965.
[9]	Guhr O, Schumann A-K, Bahrmann F, et al. Fullstop: multilingual deep models for punctuation prediction[C]//Proceedings of the Swiss Text Analytics Conference 2021. Winterthur: CEUR Workshop Proceedings, 2021.
[10]	Panayotov V, Chen G, Povey D, et al. Librispeech: an asr corpus based on public domain audio books[C]// 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). South Brisbane: IEEE Press, 2015: 5206-5210.
[11]	Hernandez F, Nguyen V, Ghannay S, et al. TED-LIUM 3: twice as much data and corpus repartition for experiments on speaker adaptation[C]// International Conference on Speech and Computer. Cham: Springer International Publishing, 2018: 198-208.
[12]	Carletta J. Unleashing the killer corpus: experiences in creating the multi-everything AMI Meeting Corpus[J]. Language Resources and Evaluation, 2007, 41 (2): 181-190. doi: 10.1007/s10579-007-9040-x URL
[13]	Li J Y. Recent advances in end-to-end automatic speech recognition[PP/OL]. V1. arXiv (2021-11-02) [2026-04-03]. https://doi.org/10.48550/arXiv.2111.01690.
[14]	Moritz N, Hori T, Le J. Streaming automatic speech recognition with the transformer model[C]// ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). Barcelona: IEEE Press, 2020: 6074-6078.
[15]	Miller R B. Response time in man-computer conversational transactions[C]// Proceedings of the December 9-11, 1968, Fall Joint Computer Conference, Part I. San Francisco: Thompson Book Company, 1968: 267-277.
[16]	Graves A. Sequence transduction with recurrent neural networks[PP/OL]. V1. arXiv (2012-11-14) [2026-04-03]. https://doi.org/10.48550/arXiv.1211.3711.
[17]	Morris A C, Maier V, Green P D. From WER and RIL to MER and WIL: improved evaluation measures for connected speech recognition[C]//Interspeech 2004. Jeju: ISCA, 2004: 2765-2768.
[18]	李鲲程, 郝航, 温仰飞, 等. 基于首Token和平均Token响应时间的大模型负载测试方法研究[J]. 通信管理与技术, 2025 (5): 6-10.
[19]	Crankshaw D, Wang X, Zhou G, et al. Clipper: A {Low-Latency} online prediction serving system[C]// 14th USENIX Symposium on Networked Systems Design and Implementation (NSDI 17). Boston: USENIX Association, 2017: 613-627.

版本	LibriSpeech WER	TEDLIUM WER	AMI WER
Topline:Faster-Whisper(Offline)	3.69%	8.66%	29.79%
A1: Faster-Whisper(基础分片)	25.65%	27.64%	73.72%
A2: Faster-Whisper(滑动窗口优化)	12.66%	12.22%	36.76%
B1: Sherpa-onnx (原生流式)	5.49%	11.82%	31.16%
B2: Sherpa-onnx + BERT(语义后处理优化)	5.49%	11.82%	31.16%

版本	LibriSpeech WER	TEDLIUM WER	AMI WER
Topline:Faster-Whisper(Offline)	3.69%	8.66%	29.79%
A1: Faster-Whisper(基础分片)	25.65%	27.64%	73.72%
A2: Faster-Whisper(滑动窗口优化)	12.66%	12.22%	36.76%
B1: Sherpa-onnx (原生流式)	5.49%	11.82%	31.16%
B2: Sherpa-onnx + BERT(语义后处理优化)	5.49%	11.82%	31.16%

版本	单线程RTF	最大安全并发数/路	满载RTF
A1: Faster-Whisper(基础分片)	0.115	9	0.975
A2: Fster-Whisper(滑动窗口优化)	0.195	6	0.981
B1: Sherpa-onnx (原生流式)	0.057	51	0.992
B2: Sherpa-onnx + BERT(语义后处理优化)	0.070	19	0.992

版本	单线程RTF	最大安全并发数/路	满载RTF
A1: Faster-Whisper(基础分片)	0.115	9	0.975
A2: Fster-Whisper(滑动窗口优化)	0.195	6	0.981
B1: Sherpa-onnx (原生流式)	0.057	51	0.992
B2: Sherpa-onnx + BERT(语义后处理优化)	0.070	19	0.992

配置	单线程 TTFT	满载 TTFT
A1: Faster-Whisper(基础分片)	1.165	2.006
A2: Faster-Whisper(滑动窗口优化)	1.369	2.403
B1: Sherpa-onnx(原生流式)	1.066	2.059
B2: Sherpa-onnx+BERT(语义后处理优化)	1.099	2.136

面向实时语音转写场景的流式与非流式语音识别性能对比研究

Performance comparison of streaming and non-streaming speech recognition for real-time transcription scenarios

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 5

参考文献 19

相关文章 15

编辑推荐

Metrics

本文评价

[1]	崔镇东, 刘笑雨, 高晖. 5G-A驱动通信与智能融合应用新范式^*[J]. 信息通信技术与政策, 2026, 52(9): 20-28.
[2]	聂山棚, 吴荻, 杜敦伟. 面向输电安全的AI终端异物检测技术与差异化部署策略研究[J]. 信息通信技术与政策, 2026, 52(8): 59-66.
[3]	黄鑫, 梁栗炎, 张倩. 运动想象脑机接口解码算法研究综述[J]. 信息通信技术与政策, 2026, 52(6): 9-17.
[4]	张云龙, 童锦瑞, 向勇, 张志强, 姚桂花, 袁博. 面向数据集建设的平台工具体系与落地实践研究[J]. 信息通信技术与政策, 2026, 52(5): 32-40.
[5]	罗敏, 傅蓉蓉, 周佳琳. AI眼镜关键技术与产业生态研究[J]. 信息通信技术与政策, 2026, 52(3): 50-57.
[6]	杨一禾, 马飞, 李哲, 董晓慧. 算力互联网赋能生成式AI视频的内容安全研究[J]. 信息通信技术与政策, 2026, 52(2): 69-77.
[7]	刘玲, 周一青, 时宁哲, 袁文昊, 石晶林. AI使能的5G-A与星闪短距融合工业无线智联网络架构与关键技术^*[J]. 信息通信技术与政策, 2025, 51(11): 2-9.
[8]	贺倩, 金伟, 徐昊天. 面向拟态计算系统的测试评估体系研究^*[J]. 信息通信技术与政策, 2025, 51(10): 40-44.
[9]	沈金辉, 宋雨辰, 田抒川, 王扬, 腊志垚, 刘宁. 先进计算驱动的非招标采购智能化应用研究[J]. 信息通信技术与政策, 2025, 51(10): 51-58.
[10]	余鹏, 陈果, 吴美琪, 罗合, 陈鸿祥. 基于数字孪生体的水电装备运维领域数据治理方法研究[J]. 信息通信技术与政策, 2025, 51(10): 65-72.
[11]	常浩伦, 臧磊, 周颖. 基于互联网教育应用软件的多维信息编码设计及应用^*[J]. 信息通信技术与政策, 2025, 51(10): 87-91.
[12]	录天凤, 陶祥忍, 景浩盟. 基于标识解析与区块链融合的园区内产品碳足迹建设路径探究[J]. 信息通信技术与政策, 2025, 51(5): 2-11.
[13]	王尧, 郭凌超, 王晨, 董颖璇, 姜文天. 人形机器人技术与标准体系研究[J]. 信息通信技术与政策, 2025, 51(4): 86-90.
[14]	刘玉婷, 吴彤, 任祥辉. 基于Nginx负载均衡的大规模脑电信号处理集群架构设计与实现[J]. 信息通信技术与政策, 2025, 51(3): 16-24.
[15]	李志芳, 成苈委, 周洁. 基于脑电图的情绪识别机器学习方法比较分析[J]. 信息通信技术与政策, 2025, 51(3): 59-67.