ufakzeka-kulak Turkish speech recognizer released
Original title: ufakzeka-kulak: Building and Evaluating a 31.6M-Parameter Turkish Speech Recognizer from Scratch
Open-source 31.6M-parameter Turkish speech-to-text model outperforms Whisper base on FLEURS and Common Voice tests.
ufakzeka-kulak is a 31.6M-parameter FastConformer encoder with token-and-duration transducer and CTC head for Turkish speech-to-text. Trained from scratch on 4,977 hours of licensed data, it outputs cased punctuated Turkish with numbers as words. With greedy decoding the normalized WER is 0.0921 on FLEURS Turkish and 0.0916 on Common Voice Turkish, lower than Whisper base and small with paired intervals wholly below zero. On conversational speech WER is 0.3807. The recipe and checkpoint were selected on development data with one-time test scoring.
- 4,977 training hours from sources allowing commercial use
- 281 hours human-transcribed
- normalized WER 0.0921 on FLEURS Turkish
- WER lower than Whisper base and small
- released under Apache-2.0
From the original article
1 min read at arxiv.orgufakzeka-kulak is an open Turkish speech-to-text model of 31. 6M parameters, a FastConformer encoder with a token-and-duration transducer (TDT) and a CTC head, trained from scratch and released under Apache-2. 0.
By Sait Furkan Teke · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.