The initial model that was used for
https://stremma.ai service, serves as a baseline for the current v2 in production, handling most of the ASR tasks.
Stremma ELM is a transformer-based encoder-decoder model, also referred to as a sequence-to-sequence model. It was trained on 0.1M hours of labeled audio and 0.4M hours of pseudolabeled audio collected via Stremma SaaS during period of work in manual transcribing.
The models were trained on either English or Danish data. The English and Danish models were trained on the task of speech recognition. The multilingual extensions were trained on both speech recognition and speech translation.
