Rohan Paul· @rohanpaul_ai · X·· 3 小时前AI 评分35
AI 导读
Encoder-Adapter-LLM 架构将"听音频"与"对转录文本进行推理"分开：encoder 提取声学信息，adapter 将这些特征桥接进 LLM，Qwen3-4B 利用语言上下文生成文本。 当语音变得口语化或杂乱时，这种语言上下文可能有帮助，但这也让弱音频幻觉和过度纠正值得仔细测试。 https://huggingface.co/HojoAI/Hojo-ASR-Multi-V1
正文
The Encoder-Adapter-LLM setup separates hearing the audio from reasoning about the transcript: the encoder extracts acoustic information, the adapter bridges those features into the LLM, and Qwen3-4B uses language context to produce the text.
That language context could help when speech gets informal or messy, but it also makes weak-audio hallucination and over-correction worth testing carefully.
来源：Rohan Paul · x.com