The technical strength of Huoshan Voice TTS has been certified by the National Inspection and Quarantine Center, with a MOS score as high as 4.64

Recently, the Volcano Engine speech synthesis product obtained the Speech Synthesis Enhanced Inspection and Testing Certificate issued by the National Speech and Image Recognition Product Quality Inspection and Testing Center (hereinafter referred to as the "AI National Inspection Center"). The basic requirements and extended requirements for speech synthesis have reached the highest level standard of the AI ​​National Inspection Center. This evaluation is conducted from the dimensions of Mandarin Chinese, multi-dialects, multi-languages, mixed languages, multi-timbres, and personalization. The product's technical support team - the Volcano Voice Team provides a rich sound library. After evaluation, its timbre MOS The highest score can reach 4.64 points, which is an industry-leading level.

As the first and only national-level voice and image product quality inspection system in the field of artificial intelligence in my country Inspection and testing agency , AI National Inspection Center has been committed to promoting the healthy development of the intelligent voice industry. Obtaining the authoritative certification from the AI ​​National Inspection Center this time also fully demonstrates that the speech synthesis technology capabilities of Huoshan Voice have reached the industry-leading level.

Feel the effect of volcano speech synthesis: ​​https://www.php. cn/link/8e0ce414531179ae9b7f60e20351ee8b##​​

More sound experience: ​​ https://www.php.cn/link/a1ada9947e0d683b4625f94c74104d73​​

For a long time, Huoshan Voice has been targeting Bytedance’s major business lines and Volcano Engine ToB industry and innovative scenarios provide industry-leading AI voice technology capabilities and excellent full-stack voice product solutions. At present, the team's speech recognition and speech synthesis cover multiple languages ​​and dialects, covering audio and video, audio reading, voice interaction, games, advertising and other application scenarios, and provide services for Douyin, Jianying, Feishu, Tomato Novels, Pico, etc. The core business provides leading voice capabilities.

It is understood that the Volcano Engine speech synthesis product

participating in this competition was independently developed by the Volcano Voice team and uses the industry's leading generative neural network technology , mainly composed of It consists of three major modules: front-end text analysis, acoustic model, and vocoder. The specific introduction is as follows:

Front-end text analysis: mainly responsible for intelligibility, such as text regularization (such as Convert numbers into year readings, number readings, etc.), character pronunciation conversion (such as Chinese phonetic notation, especially to solve the problem of multi-phonetic characters), as well as word segmentation and rhyme prediction, etc.
• Currently, the Huoshan Voice team relies on multi-task models and neural network regularization to support 12 mainstream minor languages ​​at the same time, with remarkable results.
Acoustic model: Mainly responsible for modeling from linguistic features to acoustic features. Data shows that the back-end accuracy rate of Huoshan Voice TTS can reach 99.90%. At the same time, the model can also support refined control of multiple emotions and styles, transfer of styles between different timbres, and achieve multilingual synthesis effects using only training data in a single language.
• Vocoder module: Mainly responsible for modeling acoustic features to audio signals. Now the Volcano Voice team has self-developed a vocoder based on adversarial neural network modeling, with an accuracy rate of up to 99.95%. Relying on lightweight model design and engineering optimization, the real-time rate in the cloud can reach more than a hundred times.
• The Volcano Engine speech synthesis product has a real and natural sound, vivid interpretation, and diverse styles. At the same time, it restores the rhythm of real people in a fine-grained manner and realizes various side effects such as laughter. The language phenomenon brings an immersive listening experience to people. This is the case with the supernatural dialogue speech synthesis technology recently released by the
Volcano Voice Team. Compared with traditional TTS, it can perfectly reproduce details such as modal particles, inhalation sounds, pauses during hesitation, and pronunciation prolongation, and only requires a conventional sound library. 1/4 data.

In addition, the

"tone reproduction technology" that was popular on the Internet was also developed by the Volcano Voice team. Different from the high threshold requirements of traditional speech synthesis technology for data, the Volcano voice timbre reproduction technology requires only 0.3% of the data amount of traditional methods. Ordinary people can work in a relatively quiet environment. Recording in an open environment for more than 2 minutes can meet the standards of timbre space modeling and generate an AI model of exclusive timbre, which is convenient and efficient.

Currently, Huoshan Voice will bring its speech technology capabilities that have been polished for many years to the market and open them to external companies through the Volcano engine. It has covered many application scenarios such as automobiles, finance, audio reading, video dubbing, etc., and has assisted Ruhe Many leading companies in the industry, such as Volkswagen Automotive and Zhuishu Artifact, have realized the application and expansion of AI voice capabilities. In the future, Huoshan Voice will continue to explore the efficient combination of cutting-edge technology and business scenarios, and continue to inject innovative energy into user experience and business growth. , to achieve greater value.

The above is the detailed content of The technical strength of Huoshan Voice TTS has been certified by the National Inspection and Quarantine Center, with a MOS score as high as 4.64. For more information, please follow other related articles on the PHP Chinese website!