A compact Turkish speech recognizer has outperformed Whisper base and Whisper small on two held-out read-speech benchmarks in tests conducted by its developer. The 31.6-million-parameter ufakzeka-kulak reached normalized word error rates near 9%, although Whisper large-v3 recorded lower errors on both tests. Sait Furkan Teke of ufak AI describes the model in an arXiv preprint published on November 2, 2026; peer review has not been verified.
The model was trained from scratch using NVIDIA NeMo. It combines a FastConformer audio encoder with a token-and-duration transducer (TDT), which predicts text tokens and their durations, and an auxiliary connectionist temporal classification (CTC) output. It produces Turkish text with capitalization and punctuation, writing numbers as words. The author reports releasing weights under Apache-2.0 and exports for several runtimes, intended to support offline transcription.
Training used 4,977.44 hours across 1,023,786 utterances from sources reported to permit commercial use. About 281 hours had human transcripts. The remaining 4,696 hours used uploader subtitles or machine-generated transcripts filtered through recognizer agreement. Whisper large-v3-turbo transcribed the audio; acceptable uploader subtitles were retained, while machine transcripts otherwise required agreement with Omnilingual ASR. Such filtering can still preserve errors shared by both recognizers. The author also reports removing text and audio matches with evaluation material, but incompletely disclosed training data for the labeling models and Whisper limit certainty about their prior exposure to test material.
Training, checkpoint and export choices used development data. According to the paper, a release rule fixed before training required the final model to beat Whisper base on both FLEURS Turkish and Common Voice 27.0 Turkish test sets, with paired 95% intervals entirely favoring the new model. One final model was tested once and passed. Comparisons used the same Turkish text normalization; Whisper base and small received whichever decoding setting performed better on each dataset. The main model used greedy TDT decoding in NeMo float32.
On FLEURS, comprising 743 utterances and 2.61 hours, normalized word error rate (WER) was 9.21%, with a 95% interval of 7.92–10.55%, versus 24.61% for Whisper base. On Common Voice, with 11,874 utterances and 14.54 hours, WER was 9.16% (8.73–9.58%), versus 32.87% for base. Against Whisper small, the paired WER differences were −4.21 percentage points on FLEURS (−5.06 to −3.40) and −8.85 points on Common Voice (−9.30 to −8.43). Whisper large-v3 scored 6.04% and 7.67%, respectively; no paired significance test against it was reported.
The intervals came from 2,000 bootstrap resamples, grouping by speaker where identifiers were available and otherwise by sentence or clip. They describe evaluation-sample uncertainty, not variation across training runs: the released model used one training seed. These developer-reported results have no independent replication in the supplied evidence and establish performance in the tested settings, without showing that size or architecture caused the gains.
Everyday conversation was substantially harder. Real-TurnTurk WER reached 38.07% (32.70–46.40%), roughly four times the read-speech results, on a dataset containing only six speakers. Conversational full-stop recall was 17.1% and comma recall 3.1%. MediaSpeech WER was 10.42% (10.13–10.73%), but reference formatting contributes errors, and treating clips as independent may make its interval too narrow. FLEURS also lacks speaker identifiers. Normalized WER removes punctuation, casing and fillers, so the headline scores do not capture complete transcript quality.
Audio and deployment conditions also mattered. On FLEURS, 6-kbit/s Opus compression increased WER by 5.17 percentage points (4.58–5.79). Export acceptance required development-set WER to stay within 0.5 percentage points of the corresponding NeMo decoder, but shipped sherpa-onnx paths degraded more on quiet audio. The MLX path was evaluated on a Linux CPU, without Apple-silicon testing. Only Turkish was assessed, with no demographic or regional-accent breakdown. The model requires recordings to be segmented into utterances, does not separate speakers, and has unvalidated word timestamps.
Cover illustration generated with AI; conceptual, not research evidence.