CoSTA post-hoc adapter claims ~1.7× faster spoken dialogue decoding

A new preprint on arXiv proposes CoSTA, a post-hoc adapter that reports roughly 1.7× faster speech-token generation in spoken dialogue models while leaving the original speech tokenizer and decoder untouched.

The authors describe CoSTA as "a compatibility-preserving post-hoc framework that decouples generation and synthesis token spaces while retaining the original speech tokenizer and decoder." It extends the generation vocabulary with compressed representations through mean-initialized vocabulary extension, pronunciation adaptation, and dialogue adaptation, then deterministically expands those tokens back to the original speech-token space before synthesis.

On bilingual speech-dialogue benchmarks the paper claims "approximately 1.7× expanded speech-token generation speed, lower full-response latency, and higher throughput while preserving response quality," and flags "a practical trade-off among compression, decoding efficiency, and quality." The abstract asserts cross-architecture transfer and complementarity with low-rate tokenization, but names no base models, no per-benchmark figures, and no language pairs.