Digisensus Digisensus
Research

We share our research in ASR, LLMs and other AI technologies

Digisensus Research publishes studies on speech recognition, large language models and the other technologies behind our products.

Studies

Published studies

Study ltd26

Lithuanian speech recognition: effects of dialect training and transcript spelling

Does adding dialect recordings improve Lithuanian speech recognition, and should their transcripts use dialect or standard spelling? A controlled fine-tuning study on the LIEPA-3 corpus with 18 training runs, a pre-registered protocol and an independent customer-service test.

  • Dialect WER falls from 41.62% to 31.30% with 82 h of dialect speech; other speech of the same length reaches only 40.04%
  • The first 25 hours give about two thirds of the gain
  • Customer-service WER falls from 44.89% to 39.06% with standard-spelling dialect transcripts

Released: Code (MIT), three datasets and six fine-tuned models (CC BY 4.0), and the full report as PDF.

Experiments

Experiments

Short write-ups of one question we tried and answered. Lighter than a study: no pre-registration, and a negative result is published as readily as a positive one.

Experiment

Looping transformers on the most popular open-source models: nothing got smarter

Training-free looped transformers run a few of a model’s middle layers more than once, with no retraining. We tried it in vLLM on two of the most popular open-source models, Qwen3.8-27B and Qwen3.8-Flash-Next, a 125B mixture-of-experts model. Nothing got smarter, and one scoring trap nearly fooled us.

  • Dozens of 27B configs and 34 Flash-Next configs: none beat the baseline beyond noise, and several 27B windows made it significantly worse
  • The paper’s mechanism reproduced: damping beats naive looping, the first layers cannot be looped, narrow windows are safe
  • A metric trap: raw P(gold) rewarded the model for answering with a letter, not for picking the right one
  • Qwen3.8-27B
  • Qwen3.8-Flash-Next
  • MMLU-Pro
  • GPQA-Diamond
  • vLLM

Build on it

The datasets, models and code from every study are on the Open source page, with their licenses.

Open source GitHub Hugging Face