Delivering real-time text-to-speech with Gradium

Gradium TTS is a streaming text-to-speech model for real-time voice agents that Baseten says can deliver first playable audio faster than typical human response times, reporting a 214 ms median time-to-first-audio in September 2026 Coval benchmarks. Its Delayed Streams Modeling architecture, developed by Kyutai, aligns text and audio on an 80 ms shared clock and generates speech shortly after the first text tokens arrive, with audio trailing text by a controlled number of steps to balance context, voice quality, and latency. The system uses a temporal transformer to interpret incoming text and prior audio context, alongside a smaller depth transformer that progressively generates detailed audio tokens for each frame. To reduce overhead in small, latency-sensitive models, Gradium combines GPU kernel fusion with CUDA Graphs, which it says reduced median time-to-first-audio by more than 2.5 times at small batch sizes. Baseten supports deployment scaling through its Delivery Network, which mirrors model weights, uses tiered caching, and coordinates downloads among nodes to reduce cold-start times by an estimated two to three times during traffic spikes.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.