Vikhr Salt is a multimodal model based on a pre-trained large language model, extended with new audio tokens to handle both TTS (text-to-speech) and ASR (automatic speech recognition) tasks. The model incorporates two variants for encoding audio—Encodec and SpeechTokenizer—and achieves stable training by fine-tuning precision settings. This approach allows Vikhr Salt to leverage pre-existing LLM knowledge while effectively generating and understanding speech, marking a step forward in multimodal learning.
Model Authors
Ksenya Sycheva, Konstantin Korolev, Aleksandr Nikolic
