Turkish speech recognizer with 31 million parameters beats some existing models
ufakzeka-kulak: Building and Evaluating a 31.6M-Parameter Turkish Speech Recognizer from Scratch
Computation and Language
Summary
Understanding spoken Turkish and turning it into text is hard because of the language’s unique features and limited resources. The authors built a new computer program called ufakzeka-kulak that listens to Turkish speech and writes it down with correct punctuation and capitalization. They trained it on nearly 5,000 hours of audio that could be used commercially, mixing exact and checked noisy transcripts. This new model performs better than some well-known alternatives on standard Turkish speech tests, though it makes more mistakes on casual conversations.
What this means in practice
- •For speech technology developers: Use the ufakzeka-kulak model to build Turkish speech-to-text applications with punctuation and casing without relying on external models.
- •For automatic captioning teams: Deploy the model to generate more accurate, commercially viable transcripts for Turkish video content where licensing conditions matter.$Commercial implications: Enables sale of Turkish speech transcription services with a model trained on properly licensed data for commercial use.
Authors
Sait Furkan Teke
Abstract
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, 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. 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. 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; Whisper large-v3 (greedy) scores 0.0604 and 0.0767. On conversational speech the WER is 0.3807. We also report punctuation and casing, degraded audio, and how closely each shipped export reproduces the trained model.