[论文] ufakzeka-kulak: Building and Evaluating a 31.6M-Parameter Turkish Spee...

论文概要

研究领域: NLP 作者: Sait Furkan Teke 发布时间: 2026-11-02 arXiv: 2611.03605

中文摘要

ufakzeka-kulak 是一个开放的土耳其语语音转文本模型，参数量 3160 万，采用 FastConformer 编码器 + token-and-duration transducer（TDT）+ CTC 头，从头训练，以 Apache-2.0 发布。它能输出带大小写和标点的土耳其语文本，数字以单词形式书写。其 4,977 小时训练数据均来自允许商用的来源：281 小时为人工转写，其余 4,696 小时的训练文本为上传者字幕或机器转写，仅保留「第二次独立音频阅读与之一致」的部分。配方、检查点、标注与导出选择均在开发数据上做出，测试集在训练前固定的发布门中只评一次。使用贪心 TDT 解码，归一化词错误率（WER）在 FLEURS 土耳其语测试集上为 0.0921（95% 区间 0.0792-0.1055），在 Common Voice 27.0 土耳其语测试集上为 0.0916（0.0873-0.0958），在两个测试集上均低于 Whisper base 与 Whisper small，且配对区间完全低于零；Whisper large-v3（贪心）分别为 0.0604 与 0.0767。在会话语音上 WER 为 0.3807。我们还报告了标点与大小写、退化音频，以及每个发布导出对训练模型的复现程度。

原文摘要

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 interv...

#论文 #arXiv #NLP #小凯