Aqib Nawab (@1Muhammad_AQIB) header
Aqib Nawab (@1Muhammad_AQIB) profile picture

I build voice AI agents that handle real calls: speech-to-text, LLMs, TTS, telephony and tool calling. TypeScript and Node. GSoC 2026. Open to remote roles.

Joined September 2022
Photo by Aqib Nawab (@1Muhammad_AQIB)Photo by Aqib Nawab (@1Muhammad_AQIB)Photo by Aqib Nawab (@1Muhammad_AQIB)Photo by Aqib Nawab (@1Muhammad_AQIB)Photo by Aqib Nawab (@1Muhammad_AQIB)Photo by Aqib Nawab (@1Muhammad_AQIB)Photo by Aqib Nawab (@1Muhammad_AQIB)Photo by Aqib Nawab (@1Muhammad_AQIB)Photo by Aqib Nawab (@1Muhammad_AQIB)Photo by Aqib Nawab (@1Muhammad_AQIB)
I build voice AI agents that answer real calls. On my last one, I cut response time from about 2 seconds to 800 ms and brought word error rate down from 15% to 7%. Most of that work was unglamorous tuning, and it mattered more than any model swap. I'm open to remote voice AI engineering roles.
45
agree
in 12 months there will be more compute than anyone knows what to do with. every enterprise will ask: why are we paying anthropic and openai this much, and handing them our data for free? they'll move 90% of their workloads to open models. we're already there. 90%+ of Command Code usage runs on open models, at 100T+ tokens a month. by the time vc world wakes up to the huge potential of open models, it'll already be too late. future is now. it belongs to the brave.
2
What stupid UX to show you are disconnected
8
GPT-Live-1.1 is now in OpenAI's API, and standby bills quiet minutes at $0.02 instead of $0.05. The docs say the first reply after standby may take longer, so I'd test a slow tool call that outlasts the idle timeout before trusting the savings.
16
Reson8 says that after fixing human mistakes in one well-known speech test, its new open-weights Rumble-2-Flash model's misses fell from about 6% to under 3%. Half the "errors" were in the answer key, so I'd spot-check the reference transcripts before comparing STT models.
23
Interact's new Interruption-Bench counts how often a voice model keeps talking when you say "uh-huh". Interact-0 kept going 65% of the time, and Gemini 3.8 Live only 1%. A model that never stops would score 100%, so I'd want the same test for "wait, that's wrong".
1
31
Fish Audio says Live 3 can react in as little as 186 ms, but that's the time for a short "mm-hmm", not a full answer, and they haven't published turn-taking scores yet. I'd test it with a caller saying "yeah" mid-sentence and then "wait, I need Thursday" on a noisy line.
10
Voice models need better intelligence
ChatGPT hallucinating live on stage while Sam Altman just watches No script. No safety net. Just pure confidence meeting pure fiction His reaction is everything
9
Claude Meetings hears you when you say "Claude" in a Google Meet call, but by default it answers in writing in the chat. When someone asks out loud to use a connected tool, it confirms who's speaking each time. In a group call, a voice tells you what was said, not who said it.
10
Meta's Muse can now call US businesses for you, like the cable company or the DMV. Those calls start with a phone tree and a long hold, so the hard part is telling a recorded "please stay on the line" from a real person picking up. That's the first thing I'd test.
10
Bland's Tofu-1 page describes a "TV Test": the agent asks "you want to cancel your policy?", the caller is silent, and a TV nearby says "Yes!". A transcript-only agent can cancel on that. Bland marks the example as illustrative, so I'd want real pass rates before trusting it.
13
We have so good voice models right now, but still the pass rate for the highest model is 43% 🫰
12
When a voice agent feels slow, most people blame the model. But the model is only one place where time gets lost. I went through 8 engineering changes OpenAI made to make voice AI feel like a real conversation. Here's what stood out 🧵
1
10
What I take from this as someone who builds voice agents: Measure from "user stops talking" to "first audio heard", not just model time. Watch p95, not the average. Interrupt your agent, say "mm-hm", stay silent for 2 seconds. Never leave silence while tools run.
1
2
RealtimeVAD reports speech-off a median 99 ms before Silero on 3,279 real turns, but its pause event fires about ten times as often. So I'd start the LLM call on pause, cancel it if the caller keeps talking, and commit only on end. On a quiet, noisy phone line the lead is gone.
10