Sber releases new voice AI models for global developers

Sber, a major Russian technology company, has opened up new avenues for developers by releasing advanced AI models focused on voice processing. These models, now available on GitVerse and Hugging Face, aim to accelerate innovation in areas ranging from voice assistants to accessibility tools.

The newly released GigaChat3.1-Audio-10B is a large language model designed to handle audio files and voice messages directly. Unlike traditional systems, it does not require speech to be converted into text first. This model can analyze the nuances of a speaker's voice to detect emotions, such as positive or negative sentiment, and can navigate through recordings up to three hours long. It can pinpoint specific discussion points, summarize segments, and even identify different speakers within a single recording, making it ideal for analyzing meetings and calls.

Complementing GigaChat, Sber also introduced the GigaAM Multilingual family of automatic speech recognition (ASR) models. This suite is notable for being the first open-source Russian model capable of handling multiple languages, and it reportedly offers the best speech recognition quality in Russian among available open-source options.

Anton Frolov, Senior Vice President at Sber, emphasized the naturalness of voice as a communication method and highlighted the critical role of accurate speech recognition and emotion detection in building user trust. He stated that Sber's release of these powerful tools to the global developer community is intended to foster advancements in AI-powered voice technologies, unlocking potential applications like real-time voice translation and assistive services for individuals with speech impairments. The multilingual capabilities of GigaAM also suggest international prospects for these technologies.

“

Voice is the most natural way to communicate with technology, but also the most demanding: any error in recognition or incorrectly read emotion immediately destroys trust in the assistant. Therefore, we consider audio a growth point for the market of AI assistants. By launching these models in open access, we give a powerful tool to developers and scientists. From synchronous voice translation to services for people with speech impairments – the spectrum of applications for voice technologies is very wide. And multilingualism and the ability to easily learn other languages open up international prospects for them.