An independently fine-tuned Turkish speech-recognition checkpoint from mpoyraz has accumulated close to 900,000 Hugging Face downloads. The model page packages a 300M-parameter XLS-R acoustic model trained on Turkish speech, along with a KenLM decoder built from Turkish Wikipedia text.

Developers get an inspectable alternative to hosted speech APIs, with disclosed training data, evaluation results, and hyperparameters. Its strongest results come from clean, read speech; performance drops substantially on noisier YouTube and podcast audio.

Hugging Face download counters include repeated and automated fetches, so the figure does not represent 900,000 users or production deployments. It still indicates sustained interest in an open Turkish ASR checkpoint.

A 300M-parameter core with Turkish data

The release starts from Meta's XLS-R backbone, a cross-lingual speech encoder pretrained on roughly half a million hours of unlabeled audio spanning 128 languages. Fine-tuning adapts that general-purpose representation to Turkish using two sources:

• Common Voice 7.0 Turkish: validated clips excluding the held-out test split.
• MediaSpeech: Turkish broadcast audio distributed through OpenSLR.

The acoustic network converts audio into character or token probabilities. A KenLM n-gram model then scores candidate word sequences using patterns learned from Turkish Wikipedia, helping the decoder select plausible transcriptions when the acoustic output is ambiguous.

Clean speech and web audio diverge

Word error rate measures substitutions, deletions, and insertions relative to the number of words in the reference transcript. Character error rate applies the same calculation at the character level. Lower values indicate fewer recognition errors.