dual-path-conditioning

The paper proposes Kraken, an LLM-based speech-to-speech translation model using low-bitrate vector quantization and dual-path source conditioning to enhance translation quality and preserve speaker prosody, built on Qwen3-8B.