ufakzeka-kulak: Building and Evaluating a 31.6M-Parameter Turkish Speech Recognizer from Scratch
arXiv cs.CL · 3 Nov 2026, 04:42 AEST
ufakzeka-kulak is an open Turkish speech-to-text model with 31.6 million parameters, trained from scratch. It achieves a normalized word error rate of 0.0921 on the FLEURS Turkish test, outperforming Whisper base and small models.
Key facts
- ufakzeka-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.
- Its 4,977 training hours come from sources whose licences allow commercial use; 281 hours are human-transcribed, and for the other 4,696 the training text is an uploader subtitle or a machine transcript, kept only where a second, independent reading of the audio agrees with it.
- With greedy TDT decoding the normalised word error rate (WER) is 0.0921 (95 percent interval 0.0792 to 0.1055) on FLEURS Turkish test and 0.0916 (0.0873 to 0.0958) on Common Voice 27.0 Turkish test, lower than Whisper base and Whisper small on both sets with paired intervals wholly below zero.
- On conversational speech the WER is 0.3807.
Read aloud
ufakzeka-kulak is an open Turkish speech-to-text model with thirty one point six million parameters, trained from scratch. It achieves a normalized word error rate of zero point zero nine two one on the FLEURS Turkish test, outperforming Whisper base and small models. ufakzeka-kulak is an open Turkish speech-to-text model of thirty one.6M parameters, a FastConformer encoder with a token-and-duration transducer TDT and a CTC head, trained from scratch and released under Apache-two point zero. Its four thousand nine hundred seventy seven training hours come from sources whose licences allow commercial use; two hundred eighty one hours are human-transcribed, and for the other four thousand six hundred ninety six the training text is an uploader subtitle or a machine transcript, kept only where a second, independent reading of the audio agrees with it. With greedy TDT decoding the normalised word error rate WER is zero point zero nine two one ninety five percent interval zero point zero seven nine two to zero point one zero five five on FLEURS Turkish test and zero point zero nine one six zero point zero eight seven three to zero point zero nine five eight on Common Voice twenty seven point zero Turkish test, lower than Whisper base and Whisper small on both sets with paired intervals wholly below zero. On conversational speech the WER is zero point three eight zero seven.
Sources in this cluster
- ufakzeka-kulak: Building and Evaluating a 31.6M-Parameter Turkish Speech Recognizer from Scratch
arXiv cs.CL · 3 Nov 2026, 04:42 AEST