Pulse STT has the lowest hallucination rate during silence. A caller goes quiet for eight seconds while looking for their card. Some STT models start filling that silence with words nobody said. Sometimes, those made-up words can take the entire conversation in the wrong direction. On audio with multiple silent gaps, Pulse recorded 6.85% WER, compared to 14.03% for Deepgram Nova 3 and 40.39% for ElevenLabs Scribe. Test Pulse on your own recordings with long pauses. Link in the first reply.

Lightning is the best TTS model for Hindi customer support in Voice Arena’s blind human evaluations. In the demo below, a customer has received his new card, but it has sat unused because he has not activated it. An AI voice agent calls him. The call begins in English, then switches to Hindi as soon as he responds. It patiently walks him through the steps. Those same instructions can sound like a pushy sales call or like someone genuinely trying to help. Wait for the moment he asks for a minute. That is the difference between a call that delivers instructions and one that delivers support.

AI voice agents have one problem. The voice can sound natural, but names and company names are often where they stumble first. A model that hasn’t encountered a name before may mispronounce it based on its spelling. Callers notice, and one mispronounced name can make the entire interaction feel less human. Lightning v3.1 Pro is built for names, brands and acronyms, with a pronunciation dictionary for the hardest ones. Try Lightning v3.1 Pro. Link in the first reply.

We tested Lightning v3.1 Pro against MiniMax Speech-2.8-hd in 22 languages. On Emergent TTS Evals, an open benchmark where AI judges hear two voices read the same text and pick the one they prefer, Lightning v3.1 Pro beats MiniMax in 20 of 22 languages. A voice agent should not sound worse because a caller speaks Greek instead of English. Test Lightning v3.1 Pro in the language your callers speak. Link in the first reply.

“Why are you threatening me?” A collections call turns tense. The customer is behind on an EMI, the promised payment date has passed, and he’s already angry. One wrong word can escalate the call. Every detail matters: ₹19,506, a 20 August due date, and a conversation that switches between Hindi and English. Hear how Lightning v3.1 Pro responds. The same words can sound like a threat or a plain statement of fact, depending on how they’re said. Switch to Lightning v3.1 Pro TTS today. Link in the first reply.

Clueso has raised a $10M Series A. Congratulations to the entire @clueso team on this milestone. Clueso is building agentic video creation for every team, from product demos and launches to training and internal comms. Every great product video needs a voice that can carry the story. That’s where @smallest_AI TTS comes in, bringing natural, controllable narration directly into Clueso’s video workflow. Proud to give a voice to what Clueso is building.

Smallest AI is now integrated with @AgoraIO's Conversational AI Engine. Build real-time voice agents with Pulse for speech-to-text and Lightning for text-to-speech. Configure both in your agent’s join config, and let @AgoraIO handle the real-time conversation. Integration link in the first reply.

𝗣𝘂𝗹𝘀𝗲 𝗶𝘀 𝘁𝗵𝗲 𝘄𝗼𝗿𝗹𝗱’𝘀 𝗯𝗲𝘀𝘁 𝘀𝗽𝗲𝗲𝗰𝗵 𝘁𝗼 𝘁𝗲𝘅𝘁 𝗮𝘁 𝗵𝗮𝗻𝗱𝗹𝗶𝗻𝗴 𝗳𝗮𝗿-𝗳𝗶𝗲𝗹𝗱 𝗮𝘂𝗱𝗶𝗼. Far-field audio is what a model hears when the speaker is far from the microphone. A caller doesn’t always hold the phone to their ear. They put it on speaker, get on with other things, and talk from wherever they happen to be in the room. On WildASR far-field audio, Pulse scored an 8.99% word error rate. The other models scored between 21.27% and 62.99%. If the speech to text gets the words wrong, the rest of the call can go wrong with it. That is why the far-field number is the one to check before you pick a model. These results come from our internal evaluation, so run your own far-field recordings through Pulse. Link in the first comment.