Kraken S2ST Uses Low-Bitrate VQ for Speech Translation

Kraken performs speech-to-speech translation by augmenting Qwen3-8B with low-bitrate tokens derived from single-layer vector quantization. It utilizes an Autowave-X decoder conditioned on source speech to preserve non-linguistic prosody and identity without requiring perfectly aligned training data.