中新经纬9月26日电 近日，奇富科技受邀出席了在希腊举办的国际语音通信与信号处理顶级会议——INTERSPEECH 2024，并发表了题为Qifusion-Net: Layer-adapted Stream/Non-stream Model for End-to-End Multi-Accent Speech Recognition的主旨演讲，全面展示了其在语音识别技术领域的成就，为中国语音技术走向世界、参与全球竞争树立了新的标杆。
AS ONE OF THE MOST PRESTIGIOUS ANNUAL CONFERENCES IN THE GLOBAL SPEECH SCIENCE COMMUNITY, INTERSPEECH BRINGS TOGETHER TOP SCHOLARS, RESEARCHERS, AND INDUSTRY LEADERS FROM AROUND THE WORLD TO DISCUSS THE LATEST ADVANCES, CHALLENGES, AND FUTURE TRENDS IN SPEECH TECHNOLOGY. This platform not only represents the highest academic level in the field of speech technology, but also an excellent place for the exchange and collision of new technologies and ideas, and its authority and influence are unparalleled in the industry.
Figure 1: Genius Technology gave a keynote speech at INTERSPEECH 2024
In the speech, QiFree introduced the new generation of Qifu speech recognition system "QiFree", which can support more than 20 dialects at the same time, which is the Chinese speech recognition system with the lowest word error rate in the domestic financial industry. In the comparison of KeSpeech, the authoritative test set in the field of Chinese accent and dialect speech recognition, with its deep accumulation in the field of automatic speech recognition (ASR), Qifu Technology has achieved a significant improvement in the accuracy of dialect accent classification, reaching 79.10%, far exceeding the baseline level of KeSpeech of 61.13%, which intuitively reflects the excellent performance of Qifu Technology in speech recognition accuracy. At the same time, in the key indicator of measuring the recognition error rate - CER (Character Error Rate), Qifu Technology has a score of 8.08%, far better than KeSpeech's 10.38%, demonstrating its efficiency and accuracy in the field of Chinese dialect recognition.
表1：奇富科技“QiFree”性能效果与KeSpeech Baseline对比
QiFree, the self-developed Chinese speech recognition system "QiFree" of Qifu Technology, breaks the dilemma that a single model can only recognize a specific single dialect, and through the innovative layer adaptive fusion structure, with the help of the shared information coding module to extract dialect information more efficiently, it realizes the speech and translation, and further enhances the real-time interaction ability of the voice robot. It is worth mentioning that "QiFree" not only maintains a leading position in the CER in the field of Mandarin recognition, but also achieves a significant improvement of more than 15% in the recognition performance of multiple dialect regions such as Hebei-Shandong, Jianghuai, Jiao-Liao, Lan-Yin, etc. This breakthrough achievement was highly recognized by three independent reviewers of INTERSPEECH, who unanimously recognized the outstanding performance of the system framework innovation and recognition performance demonstrated in the paper "Qifusion-Net: A Streaming/Non-Streaming End-to-End Multi-Accent Speech Recognition Framework Based on Feature Fusion", and unanimously awarded it an "ACCEPT" rating.
It is worth mentioning that in comparison with first-class domestic companies (such as a technology giant and the most influential open source community of speech recognition in China), Qifu Technology also shows an overwhelming advantage. Even in the face of opponents with larger parameter scale and richer training data, Chiefo Technology was still able to stand out with a lower CER (8.08% vs 15.61% vs 26.55%), proving the superiority of its technical architecture and the efficiency of algorithm optimization. In addition, compared with the world's leading speech recognition systems (such as Openai-whisper v2), although the latter has significant advantages in general language recognition, Qifu Technology still maintains a significant advantage in the field of Chinese dialect recognition, which further confirms its global leading position in dialect recognition technology.
Table 2: Comparison of the key indicators of QiFree Technology with first-class technology companies at home and abroad
OUR WONDERFUL APPEARANCE AT INTERSPEECH 2024 IS NOT ONLY A COMPREHENSIVE DEMONSTRATION OF ITS YEARS OF HARD WORK IN THE FIELD OF SPEECH RECOGNITION TECHNOLOGY, BUT ALSO A DECLARATION TO THE WORLD OF THE STRONG COMPETITIVENESS AND UNLIMITED POTENTIAL OF CHINESE ENTERPRISES IN THIS FIELD. With its outstanding technical strength and innovative spirit, Qifu Technology is leading a new round of development trend of dialect recognition technology, contributing Chinese wisdom and strength to the progress of global voice communication and signal processing technology. (Sino-Singapore Jingwei APP)