ufakzeka-kulak：从零构建并评估 31.6M 参数土耳其语语音识别器
ufakzeka-kulak: Building and Evaluating a 31.6M-Parameter Turkish Speech Recognizer from Scratch
Sait Furkan Teke (ufak AI)
中文摘要
ufakzeka-kulak 是一个 31.6M 参数的开源土耳其语语音转文本模型，采用 FastConformer 编码器，搭配 token-and-duration transducer（TDT）与 CTC 头，从零训练并以 Apache-2.0 发布。输出带大小写与标点的土耳其语，数字写作单词。训练数据 4,977 小时均来自允许商用的来源；其中 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，仅在音频的第二次独立读取与其一致时保留。配方、检查点、标注与导出选择均在开发数据上确定，测试集在训练前固定的发布门禁中仅评分一次。贪心 TDT 解码下，FLEURS 土耳其语测试的归一化词错误率（WER）为 0.0921（95% 区间 0.0792 至 0.1055），Common Voice 27.0 土耳其语测试为 0.0916（0.0873 至 0.0958），在两个测试集上均低于 Whisper base 与 Whisper small，配对区间完全低于零；Whisper large-v3（贪心）分别为 0.0604 与 0.0767。对话语音的 WER 为 0.3807。还报告了标点与大小写、退化音频，以及各发布导出与训练模型的接近程度。
关键要点
- 0131.6M 参数 FastConformer+TDT/CTC 土耳其语模型，从零训练，Apache-2.0 开源
- 024,977 小时商用许可数据，仅 281 小时人工转写，其余经二次独立校验
- 03FLEURS 与 Common Voice 27.0 测试 WER 约 0.092，低于 Whisper base/small
- 04对话语音 WER 0.3807，明显高于朗读测试集
- 05测试集在训练前固定的发布门禁中仅评分一次，避免调参泄漏
解读
尚无解读。
原始英文摘要
arXiv:2611.03605v1 Announce Type: new 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.