Gradium's latest streaming TTS hits ~50ms time-to-first-audio while improving naturalness and hard cases like phone numbers—freeing latency budget for LLM turns and barge-in in voice agents.
Why we built the fastest robust TTS model
Our latest Text-to-Speech model takes around 50ms for the first audio chunk to arrive. These are the lowest latency figures out there among frontier TTS models.
Generally latency and accuracy are traded values. ElevenLabs, for example, ships its newest model in two variants: Eleven v4, "tuned for produced content where quality matters most", and Eleven v4 Turbo, a separate variant focused on lower latency. Choosing a fast model does not have to come with cutting corners for model robustness. Gradium's latest Text-to-Speech model beats Eleven v4 Turbo on latency while improving on naturalness and expressivity in all languages, and retaining our strengths of handling hard cases like phone numbers, alpha-numeric contextual text.
Does it matter when the latency is lower than the human noticeable threshold?
The accepted latency budget voice agents or any system that talks to humans have is around 200–300ms. When a person is speaking, you consider it as their turn. When they are done speaking, it is the turn of the other person to respond or talk. Between this lies a turn gap. The latency budget is for this gap.
By reducing our latency by 170ms, we give that time back to the rest of the cascaded system: more of the budget is left for intelligence in the LLM and for handling barge-in attempts from other speakers. This improves the overall experience for your voice agents, especially for phrases that are short, snappy turns.
How businesses will gain an edge over competitors?
As the model is served realtime and streaming, applications that need to respond instantaneously to their users come across as more natural, humanlike and always present.
Live streaming avatars built as coaches, support agents that are expected to respond with solutions in no time.
Perceived latency for telephonic calls has always had room for improvement. Because alongside the latency from the voice model, there is big chunk of wait added by the carrier network. Now the model latency can be largely subtracted from this equation.
Audiobooks, narration, pre-rendered IVR prompts and any flow where the full text exists before playback starts will still gain from the model's improved expressivity compared to the previous model.
Developers can also experiment with longer first phrases: give the TTS model more context for the opening words, and it still starts speaking earlier than the previous model did.
How to use the model?
The model is now out of beta and is the default model for API and studio. If you're already using the API, no changes are required from your side.