Topic
Conversational AI systems that speak with users in real time, including their speech recognition, turn-taking and response generation stacks.
A dev.to author's nightly check of Pipecat's 67 community plugins found 19 that fail to install or import against release 1.12.0. The post says nothing checked the plugins against new Pipecat releases before, so a builder choosing a voice vendor should run the install first.
Reality
Modulate raised $25 million to put models that analyze raw call audio, not transcripts, in front of more developers. Its pitch to teams running voice agents is a separate layer that spots cloned voices and grades how agents handle callers.
Perspective Coverage
3 publishersReality
Google added lip-syncing video avatars to its Gemini 3.8 Live voice agents, available now to Gemini Enterprise customers in 97 languages. Early reviews called the faces creepy and the launch came with no customer evidence, so support teams will have to measure trust in their own queues.
Perspective Coverage
7 publishersReality
Google's two new voice models process speech and reasoning at the same time and can fire tool calls without pausing the conversation. Published scores put the stronger one top of a speech-to-speech index and lowest on banking tasks.
Perspective Coverage
3 publishersReality
In a 30-call test suite for a home services intake agent, guardrail G4 told the model to stop collecting fields the moment it heard a gas smell and never told it when to resume. The dispatcher got the result.
Reality
Oathra scores an agent phone call with separate code that reads the callee's turns, and its maintainer spent this week fixing the two natural English confirmation shapes that the checker itself got wrong.
Reality
Microsoft treats speech as its own agent kind, with a managed realtime orchestrator holding the socket open. A dev.to walkthrough sets out why turn detection and tool calls had to be restructured around a few hundred milliseconds.
Reality
Across three clinics and a bit over 1,000 real conversations, that overflow rule makes the agent's own call log the first count of ring-outs these practices have had. It is also the only source for the 20 to 30 recovered hours a month.
Reality
Reception bundles the voice model, a phone number and a calendar write into a self-serve plan with 75 included minutes a month. That works out at about 39 cents a minute, the rate a home-built inbound agent has to beat.
Reality
When the acoustic path changes mid-playback, the browser's echo canceller stops cancelling and the leftover sound reaches the voice detector as speech. The fix was 300 extra milliseconds, applied only during playback.
Reality
A voice avatar kept its microphone button lit while the server logged 0.00 volume, and the fix needed three failed hypotheses before a log showed the widget enabling a track that was about to be discarded.
Reality
Fangchang Ma says the lag and the dead-eyed listening in voice products come from the handoffs between four models. Nuance's evidence is a demo it calls unfinished and a research preview promised for later this year.
Reality
A client's request for an expressive voice pushed a team off VAPI and onto Twilio, Deepgram and Cartesia. The metric they chased was time-to-first-audio: how fast the first syllable arrives. Owning the pipeline means owning barge-in.
Reality
Measured per call rather than averaged, one voice agent's outbound cache hit rate came in at less than half of inbound on identical code and prompts, and the cause was two variable fields sitting inside the cached prefix.
Reality
Partial recognition results and speech the user never heard end up in the prompt that drives the next turn. A dev.to tutorial's fix is a reducer with exactly two commit points, and the cost is an adapter you write yourself.
Reality
AWS says three integration decisions carry Natera's phlebotomy booking agent. The reported numbers are good, and they cover a narrower job than "booked appointment".
Reality
AWS puts a Claude-powered Connect agent on a real phone number and reaches the restaurant backend over MCP. The unforgiving part is the pause the caller hears while that happens.
Reality
A dev.to breakdown makes the case that turn detection, endpointing and barge-in are the hard parts of a voice product, and a model swap does not touch any of them.
Reality
A practitioner's account argues live transfer, structured message and flat refusal are separate code paths, fired by four different triggers. That is a twelve-cell table, not a toggle.
Reality
A builder's account decomposes a production voice minute into speech-to-text, tokens, speech synthesis, telephony and media infrastructure, then puts failed calls in the numerator.
Reality