T-Tech Releases T-one for Russian Speech Recognition

T-Tech, the AI research arm of Russian digital bank Tinkoff, has released T-one, a new model for automatic speech recognition (ASR). The model is specifically designed for transcribing Russian-language audio from telephone calls, a common but challenging domain for speech-to-text systems.

The model uses a streaming Conformer architecture, which allows it to process audio in real-time with low latency. This makes it well-suited for live transcription applications like call center monitoring or voice assistant interactions. According to its creators, T-one was trained on a substantial dataset of over 30,000 hours of Russian speech to achieve its specialized performance.

Use Cases and Licensing

Given its focus on telephony, T-one is primarily aimed at enterprise and research applications involving voice data. The model's architecture and training are tailored to handle the lower audio quality and specific conversational patterns found in phone calls.

While the model weights are publicly available on the Hugging Face Hub, they are released under a Creative Commons Non-Commercial (CC BY-NC 4.0) license. This permits use for research and experimentation but restricts deployment in commercial products without a separate agreement with T-Tech.