voicekit-team/t-one

T-one is a tool that converts spoken Russian audio into text in real time, designed specifically for phone call transcription. If you have a call center recording or a live phone conversation, T-one listens to the audio and produces a written transcript as the speech happens, rather than waiting until the recording ends. Under the hood, it processes audio in small 300 millisecond chunks, continuously feeding each chunk into a Conformer-based acoustic model, a neural network architecture designed for speech. That model generates probabilities for each letter in the Russian alphabet, frame by frame, while keeping track of context from previous chunks so words are not cut off at chunk boundaries. A splitter component then watches those letter probabilities to detect where phrases begin and end, based on gaps of silence between speech. Finally, a decoder converts those letter probabilities into actual words, either by picking the most likely letter at each step, called greedy decoding, or by using a separate language model to pick better sounding word sequences, called beam search decoding. The README includes benchmark comparisons against other Russian speech recognition models, measured by word error rate, the percentage of words the system gets wrong compared to a correct transcript. T-one is shown scoring lower error rates than several larger competing models on call center audio, despite having fewer parameters itself. You would use T-one if you need to transcribe Russian phone calls, build a voice bot, or add live captions to a Russian language audio stream. It ships as a Python library and also as a Docker container with a browser based demo so you can test it immediately by uploading a file or speaking into your microphone. The model runs on CPU and requires a machine with at least 4 cores and 8 GB of RAM.