Are you speaking my languages? On spoken language adherence in multimodal LLMs

Large language model-based automatic speech recognition systems frequently misidentify the output language, a flaw that degrades transcription accuracy and downstream application reliability, according to a paper submitted to arXiv on 15 Jun 2026 [1]. The study, titled "Are you speaking my languages? On spoken language adherence in multimodal LLMs," formally defines the problem as a lack of language adherence and introduces a novel metric to quantify violations [1]. While LLM-based ASR enables seamless multilingual use, the models' tendency to output text in an incorrect language compromises fidelity [1]. To address this, the authors propose a soft prompting approach that hints at potential spoken languages without strictly constraining the output, preserving flexibility and code-switching capabilities [1]. The research evaluates three mitigation strategies: zero-shot prompting for robust guidance under uncertainty, supervised fine-tuning to improve prompt adherence, and Chain-of-Thought reasoning to enforce adherence during decoding [1]. A comparative analysis across multiple languages assesses each method's effectiveness in reducing language violations while maintaining overall ASR performance [1]. The paper also discusses trade-offs to guide strategy selection under various compute constraints [1]. Large language models are a type of machine learning model designed for natural language processing tasks such as language generation, trained with self-supervised learning on vast amounts of text [5]. The field has seen rapid development from companies such as DeepSeek, a Chinese AI firm founded in July 2023 that launched its DeepSeek-R1 model in January 2025 [4]. DeepSeek's models are described as open-weight, meaning the exact parameters are openly shared, though the training data is not openly licensed [4]. The company reportedly reduced training expenses by incorporating techniques such as mixture of experts layers [4]. Alibaba Cloud's Qwen family of large language models is another example, with many models distributed under the free and open-source Apache 2.0 license [6]. A separate scoping review of generative systems for quantum circuit and code generation, identified through Hugging Face and arXiv, found that while all reviewed systems address syntactic validity and most address semantics, none reports end-to-end evaluation on quantum hardware [3]. This gap between generated circuits and practical deployment mirrors the broader challenge of moving from model capability to real-world reliability, a concern echoed in the ASR language adherence findings [1][3].