{"id":10262,"date":"2026-02-14T05:56:20","date_gmt":"2026-02-13T19:56:20","guid":{"rendered":"https:\/\/www.paradisec.org.au\/blog\/?p=10262"},"modified":"2026-02-14T05:57:11","modified_gmt":"2026-02-13T19:57:11","slug":"speech-recognition-on-your-laptop","status":"publish","type":"post","link":"https:\/\/www.paradisec.org.au\/blog\/2026\/02\/speech-recognition-on-your-laptop\/","title":{"rendered":"Speech recognition on your laptop"},"content":{"rendered":"\n

Following on from the work reported on two posts ago <\/a>with progress on speech recognition for Bislama and Nafsan, Aso Mahmudi has now created a desktop app (called Easper – E<\/strong>lan A<\/strong>utomated Spe<\/strong>ech R<\/strong>ecognition) that takes a wav file as input, segments it, does speaker diarisation, and transcribes it, delivering an Elan file as the output. All this is done on a laptop computer and needs no internet connection.<\/p>\n\n\n\n

As can be seen in the image below, you can set the number of speakers, and can alter the silence recognition (‘gap between segments’) and the minimum segment length, depending on the characteristics of the recording and the rate of speech. You select the language model (which has to have been created in advance) and then run the process. For example, with a 44 minute file the first pass of segmentation took less than three minutes. Transcription took seven minutes on my laptop (MacBookPro 2021, 16Gb RAM, System 26.3). It always needs checking and correcting, but is remarkably good (as reported earlier<\/a>, around 10% character error rate for Nafsan<\/a>).<\/p>\n\n\n\n

\"\"<\/a><\/figure>\n\n\n\n

Easper also provides Elan file analysis, as seen below, with the following results given:<\/p>\n\n\n\n