Gemma 4 语音编码器冻结权重下的沙特与海湾阿拉伯语识别
原标题：How Far Can the Gemma 4 Speech Encoder Go Without Changing Its Weights? Saudi and Gulf Arabic Speech Recognition with a Frozen and an Adapter-Augmented Encoder
AI 摘要
该研究探索在不改变 Gemma 4 共享的 305M 参数语音编码器权重的前提下，将其用于沙特与海湾阿拉伯语语音识别。冻结系统仅训练 22.6M 参数的 CTC 读取器，适配系统额外加入 3.1M 低秩适配器与 41.5M 混合 CTC/注意力读取器。结果显示冻结系统在字符错误率上与 2.07B 参数的 Cohere Transcribe Arabic 在六项人工标注条件中的五项无统计差异，且 CPU 上快 3.4 倍；适配系统在部分测试集上 WER 与 Cohere 无显著差异甚至更优，但在长篇沙特语音上仍落后。
正文节选
How Far Can the Gemma 4 Speech Encoder Go Without Changing Its Weights? Saudi and Gulf Arabic Speech Recognition with a Frozen and an Adapter-Augmented Encoder Abstract We study the 305M-parameter audio encoder shared by Gemma 4 and EmbeddingGemma 2 (identical except for its final projection) for Saudi and Gulf Arabic speech recognition, without changing any of its weights. A frozen system trains only a 22.6M-parameter CTC reader over a mix of encoder layers; an adapted system adds 3.1M low-rank