“Wait, that’s not what I meant 🙄” Our CTO Carter Huffman talks to an AI voice assistant every day. When he says that, it keeps talking for about four more seconds. A human would stop instantly. Or get called rude. That’s where voice AI is right now: useful, but still young. It fails in ways people don’t expect and don’t know how to design around. One day it’s perfect. The next it’s mispronouncing words, stuttering, lagging, talking over the caller. Most of that never shows up in a transcript. The words look fine. The conversation wasn’t. Carter joined The Microdose on to break down why, when you’re running voice agents across thousands of calls, you can’t read your way to these failures. You have to listen ↓ That’s what Modulate’s Velma is built for: monitoring AI voice agent conversations in real time, hearing what the transcript leaves out, and flagging issues to your team or your AI before they escalate

Good comms travel. In Call of Duty®: Warzone™, ~5.5% of players produced at least one prosocial voice interaction. ~22% were exposed to one. That’s roughly 1 speaker for every 4 players reached. Modulate’s new case study looks at prosocial voice in North American Warzone™ matches from Dec 4, 2025 to Feb 4, 2026, identified by Modulate’s ToxMod, powered by Velma. What we found: → ~66% of matches with prosocial voice also had a cooperative gameplay event, like a revive or a thanks → Same-session continuation: ~77% for prosocial speakers, ~71% for exposed players → Exposure reached active, new and returning players, not just the most engaged Healthy communities aren’t defined only by reducing harm. Positive play is already happening at meaningful scale, and it’s worth understanding. @Activision

Your next inbound call might not be a person. Part 2 of our LinkedIn Live series: Carter Huffman and William Brown go under the hood of Modulate’s Synthetic Voice Detection. Live demo: real-time detection routing bots and humans where they belong. October 27, 3PM ET. Save your spot ↓

“I’m fine.” The transcript says one thing. The voice can say something completely different. Velma by Modulate hears what speech-to-text misses: tone, emotion, hesitation and conversational context. Stop transcribing. Start truly listening. Build it into your own stack with Modulate APIs. 1,000 free credits to start.👇

primary_verdict: “ai-vocal-music” Modulate AI Music Detection returns a clip verdict, separate AI scores for vocals and instrumentals, and per-window results every 4 seconds.

Arabic. English. Arabic again. When a conversation switches languages, Modulate Transcribe detects the language of every utterance, so each turn comes back in the language actually spoken. → 100 languages, Arabic included → Speaker labels that hold across language switches → Batch and streaming

Clinical audiology sessions are messy 👀 Overlapping speakers. Emotional nuance flat transcripts miss. Notes that need to be instant, not "eventually." One developer needed transcription that could keep up. So they built their AI Scribe SaaS on Velma 2. The result: 👉 Live diarization 👉 Emotional context per utterance 👉 Clinical notes generated in real time Their words, not ours: ⭐ 5/5: "Accurate Transcriptions with Emotional Context at a Fantastic Price." This is what audio-native voice intelligence looks like in production 💯

Same words. Five different meanings. “I’m fine” can be relief, exhaustion, sarcasm, or a cry for help. Your transcript can’t tell the difference. Velma can. Audio-native voice intelligence, in real time. Built for teams that need to know what was meant, not just what was said.