ufakzeka-kulak
A lightweight open-source Turkish speech recognition model based on a FastConformer encoder and TDT/CTC decoders, supporting output with punctuation and capitalization.
Timeline (last 90 days)
ufakzeka-kulak is an open-source Turkish speech-to-text model with only 31.6M parameters, using a FastConformer encoder and TDT and CTC heads, trained from scratch, and released under the Apache-2.0 license.
The model outputs Turkish text with capitalization and punctuation, with numbers presented as words.
The model used 4,977 hours of training data, all sources of which permit commercial use, of which 281 hours are human-transcribed and the training text for 4,696 hours comes from uploader subtitles or machine transcription.
For the 4,696 hours of non-human-transcribed data, a sample is retained only when a second independent reading of the audio matches the text.
With greedy TDT decoding, the model achieves a normalized WER of 0.0921 on the FLEURS Turkish test set, with a 95% confidence interval of 0.0792 to 0.1055.
On both the FLEURS and Common Voice datasets, ufakzeka-kulak's WER is lower than Whisper base and Whisper small, and the paired confidence intervals are entirely below zero, so the performance advantage is statistically significant.
In conversational speech scenarios, ufakzeka-kulak's WER rises to 0.3807.
All Facts (7)
ufakzeka-kulak is an open-source Turkish speech-to-text model with only 31.6M parameters, using a FastConformer encoder and TDT and CTC heads, trained from scratch, and released under the Apache-2.0 license.
50%UnverifiedThe model outputs Turkish text with capitalization and punctuation, with numbers presented as words.
50%UnverifiedThe model used 4,977 hours of training data, all sources of which permit commercial use, of which 281 hours are human-transcribed and the training text for 4,696 hours comes from uploader subtitles or machine transcription.
50%UnverifiedFor the 4,696 hours of non-human-transcribed data, a sample is retained only when a second independent reading of the audio matches the text.
50%UnverifiedWith greedy TDT decoding, the model achieves a normalized WER of 0.0921 on the FLEURS Turkish test set, with a 95% confidence interval of 0.0792 to 0.1055.
50%UnverifiedOn both the FLEURS and Common Voice datasets, ufakzeka-kulak's WER is lower than Whisper base and Whisper small, and the paired confidence intervals are entirely below zero, so the performance advantage is statistically significant.
50%UnverifiedIn conversational speech scenarios, ufakzeka-kulak's WER rises to 0.3807.
50%