Overview
- Field: NLP (speech recognition)
- Author: Sait Furkan Teke
- Date: 2026-11-02
- arXiv: 2611.03605
Summary
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. It writes cased, punctuated Turkish with numbers as words.
Its 4,977 training hours come from sources whose licences allow commercial use. 281 hours are human-transcribed; for the other 4,696 hours the training text is an uploader subtitle or a machine transcript, kept only where a second, independent reading of the audio agrees with it. The recipe, checkpoint, labelling and export choices were made on development data, and the test sets were scored once, in a release gate fixed before training.
Results
With greedy TDT decoding, the normalized word error rate (WER) is:
| Test set | WER | 95% interval | |---|---|---| | FLEURS Turkish | 0.0921 | 0.0792–0.1055 | | Common Voice 27.0 Turkish | 0.0916 | 0.0873–0.0958 | | Conversational speech | 0.3807 | — |
The model outperforms Whisper base and Whisper small on both test sets, with paired intervals entirely below zero. Whisper large-v3 (greedy) achieves 0.0604 on FLEURS and 0.0767 on Common Voice.
The paper additionally reports punctuation and casing performance, results on degraded audio, and how faithfully each released export reproduces the trained model.
*Auto-collected on 2026-11-04.*