Fixing ASR

At Amerandish, one of our key products is a Persian ASR system used by both individuals and businesses. Some business users, such as those in the pharmaceutical industry, required domain-specific ASR, especially for complex Persian medicine names.

At the time, our ASR system was based on DeepSpeech2, which incorporates an acoustic model and a language model (LM). We fine-tuned the LM to improve its predictions within specific domains by scraping relevant texts and retraining the LM with domain-specific data.

Additionally, we retrained the original LM after identifying output weaknesses. By eliminating typos in the text dataset and optimizing parameters in kenlm, we managed to improve the word error rate by 20%.