Qwen vs. Gemma Integration with Whisper: A Comparative Study in Multilingual SpeechLLM Systems
This addresses multilingual speech recognition for the MLC-SLM Challenge 2025, presenting an incremental improvement with competitive results.
This paper tackled multilingual speech recognition by combining a fine-tuned Whisper encoder with projector architectures and LLM decoders, achieving competitive performance with average WER/CER results of 16.63% using Gemma3-12B and 18.6% using Qwen2.5-7B.
This paper presents our system for the MLC-SLM Challenge 2025, focusing on multilingual speech recognition and language modeling with large language models (LLMs). Our approach combines a fine-tuned Whisper-large-v3 encoder with efficient projector architectures and various decoder configurations. We employ a three-stage training methodology that progressively optimizes the encoder, projector, and LLM components. Our system achieves competitive performance with a private test average WER/CER result of 16.63% using the Gemma3-12B and 18.6% using the Qwen2.5-7B as decoder-only language model.