# Delayed Streams Modeling

Gradium TTS pushes time-to-first-audio below typical human speech response, reaching a median TTFA of 214 ms on the Coval benchmark as of September 2026 with an interquartile range tight enough for developers to set reliable latency budgets for the rest of the pipeline. It uses Kyutai's Delayed Streams Modeling, which treats text and audio as two parallel streams on a shared clock so audio can start before the upstream model finishes the sentence. On Baseten it ships as a dedicated deployment whose weights are pre-mirrored and cached, cutting cold starts by 2–3x.

Gradium TTS pushes time-to-first-audio below typical human speech response, reaching a median TTFA of 214 ms on the Coval benchmark as of September 2026 with an interquartile range tight enough for developers to set reliable latency budgets for the rest of the pipeline. It uses Kyutai's Delayed Streams Modeling, which treats text and audio as two parallel streams on a shared clock so audio can start before the upstream model finishes the sentence. On Baseten it ships as a dedicated deployment whose weights are pre-mirrored and cached, cutting cold starts by 2–3x.

The work first shows that the padding token can be removed from Delayed Streams Modeling (DSM) while maintaining competitive recognition performance, then proposes Factorized DSM (F-DSM), which separates the waiting probability from the distribution over the original LLM vocabulary, removing ASR-specific tokens from the text prediction space and allowing the large-vocabulary softmax to be skipped on waiting steps; experiments on the Corpus of Spontaneous Japanese and LibriSpeech show that F-DSM achieves better recognition performance than DSM, greatly reduces GPU memory use while maintaining similar training throughput, provides a small inference speed improvement through softmax skipping, and reduces the degradation in text-only perplexity observed with DSM.

The work first shows that the padding token can be removed from Delayed Streams Modeling (DSM) while maintaining competitive recognition performance, then proposes Factorized DSM (F-DSM), which separates the waiting probability from the distribution over the original LLM vocabulary, removing ASR-specific tokens from the text prediction space and allowing the large-vocabulary softmax to be skipped on waiting steps; experiments on the Corpus of Spontaneous Japanese and LibriSpeech show that F-DSM achieves better recognition performance than DSM, greatly reduces GPU memory use while maintaining similar training throughput, provides a small inference speed improvement through softmax skipping, and reduces the degradation in text-only perplexity observed with DSM.