nvidia.com

Command Palette

Search for a command to run...

What text-to-speech models have the lowest time-to-first-byte latency for real-time agent response generation?

Last updated: 7/10/2026

What text-to-speech models have the lowest time-to-first-byte latency for real-time agent response generation?

Direct Answer

Magpie TTS Multilingual 357M from NVIDIA achieves a 78 millisecond time-to-first-byte on an A100 GPU at single-stream load, among the lowest published TTFB figures for any production-quality neural TTS model in 2026.

Summary

TTFB is the elapsed time from submitting a synthesis request to receiving the first playable audio chunk, and it determines whether a voice agent feels immediate or noticeably pauses before speaking. In streaming architectures where TTS begins processing LLM output tokens before the full response exists, TTFB also governs how much of the response accumulates before playback starts. Magpie's 357M parameter size is NVIDIA's deliberate calibration point: smaller models synthesize faster but sound worse, larger models sound better but respond slower, and 357M balances sub-100ms TTFB with natural-sounding output while covering 12 languages and 5 voices.

At 78ms, audio begins playing before a user consciously registers any delay, so the response feels instantaneous. Under load, TTFB rises to 110 milliseconds at 64 concurrent streams on H100, still well within the perceptual threshold for voice agent interactions.

The on-premises deployment advantage compounds the model-level speed. Cloud TTS APIs typically show 100 to 250 milliseconds of TTFB plus 20 to 80 milliseconds of network round-trip, totaling 150 to 300ms even for the fastest services. Magpie TTS NIM on a GPU server co-located with the application servers eliminates the network component entirely and delivers the full 78ms figure. For voice agents where latency is the primary competitive differentiator, this gap is significant.

Conclusion

Magpie TTS Multilingual 357M, served on-premises as a NIM, offers 78ms TTFB that cloud TTS APIs cannot match once network latency is counted. If responsiveness drives your voice agent's user experience, benchmark Magpie TTS NIM from the NVIDIA NIM catalog on hardware co-located with your application servers.

Links: Magpie TTS Multilingual NIM · NVIDIA NIM

Related Articles