Pulse: Fast to the First Word, Fast to the Last One
We ran Smallest AI's Pulse through the Speko harness and wired it into the gateway. It's the rare streaming STT that's fast on both clocks a live agent actually feels — ~64 ms to the first partial, ~180 ms to the final — while holding 5.1% WER on English.
Almost every speech-to-text leaderboard measures the same thing: word error rate on clean, pre-recorded audio. It’s a useful number, and it’s the wrong one for a voice agent. A caller never experiences your STT’s accuracy — they experience its latency. The half-second of dead air after they stop talking. The beat too long before the agent starts to understand. That gap is where a conversation stops feeling like a conversation.
So when we evaluate an STT for live agents, we watch two clocks as closely as accuracy: how fast the first words come back, and how fast the turn actually ends. Most models are good at one or neither. Smallest AI’s Pulse is the rare one that’s excellent at both.
The two clocks
| Metric | Pulse |
|---|---|
| First partial token | ~64 ms |
| End-of-turn, with end-of-speech finalization | ~180 ms |
| WER — English (FLEURS, n=50, via the Speko gateway) | 5.1% |
Fast where it’s felt
Pulse is streaming-first by design, and it shows. First partial tokens land in about 64 milliseconds — fast enough that the transcript is already scrolling while the caller is still mid-sentence. For a downstream LLM that needs to start reasoning the moment a turn is recognizable, that head start is everything.
The bigger surprise is the end of the turn. Paired with end-of-speech finalization, Pulse commits its final transcript in around 180 milliseconds after the caller stops — not the one-to-three seconds you get when an STT waits out its own silence timer. In a live call that difference is the feel of the agent: a reply that lands on the beat instead of after an awkward pause.
Fast to the first word, fast to the last one. That’s the combination real-time voice has been missing.
And the accuracy holds
Speed only counts if the words are right. On our harness Pulse turned in 5.1% WER on English (FLEURS, n=50, measured through the Speko gateway) — in the competitive band with the established names, and comfortably good enough for the tasks voice agents actually run: intake, scheduling, support, confirmations. It’s a genuinely strong accuracy-for-speed trade, and at Pulse’s price point it’s hard to argue with.
One thing on our wishlist for the next release: Pulse writes numbers and dates out as words (“eight hundred” rather than “800”), and a digits / inverse-text-normalization mode would save every agent builder a normalization pass.
Now on Speko
Pulse is live in the Speko gateway today as a routable STT — any voice agent on the platform can select it with a single config change, no separate integration, no extra key. If you’re building agents where responsiveness is the product, it’s one of the first models we’d point you at.