Twilio's Conversation Relay now ships a turn detection model and automatic provider failover. Smart Turn Detection reached general availability on November 13, 2026, and speech provider failover followed on November 18. Twilio's own changelog says the turn model "evaluates speech meaning in real time to accurately detect when a caller has actually finished speaking," replacing rigid silence timers. We read Twilio's blog and both changelog entries on November 21, 2026, and every claim below links to that source.
Key Takeaways
- Twilio announced Smart Turn Detection for Conversation Relay as generally available on November 13, 2026, as an in-house model that reads meaning instead of counting seconds of silence.
- Conversation Relay failover lets a call switch once to a backup speech-to-text provider and once to a backup text-to-speech provider, and the call ends if the backup also fails.
- Turn detection works with every supported speech-to-text provider except Deepgram Flux, which uses its own built-in turn detection.
- Twilio's announcement gives no measured accuracy for turn detection or switch time for failover, so test both on your own call audio.
- Topcalls AI calls start from $0.25 a minute, depending on the country called, and a dedicated team configures each agent.
What did Twilio announce for Conversation Relay?
Twilio announced five Conversation Relay updates in a November 17, 2026 blog post: Smart Turn Detection, automated speech-to-text failover, dynamic configuration parameters, HD Audio and an advanced conversational voice. The post marks all five as generally available. It also expands Twilio Agent Connect and Conversation Intelligence, covered below.
Conversation Relay is Twilio's layer between a phone call and an AI agent. The agent can be any model, speech provider or framework. Twilio says it is "not trying to be your AI agent or your model" and supplies the communications infrastructure underneath.

That position matters for buyers. Conversation Relay is infrastructure for teams that assemble a voice agent from parts. A team that wants a working campaign without assembling parts is a different buyer, and the questions in the last section apply to both.
How does Smart Turn Detection stop AI from interrupting callers?
Smart Turn Detection is a model that decides whether a caller has finished speaking or is only pausing. Twilio's changelog says it separates a caller who is done from one "pausing mid-sentence to think or use filler words," which removes the cut-offs that come from fixed silence timers.
Three details from the changelog are worth knowing:
- It runs in Twilio's media stack, so it works across all supported speech-to-text providers, except Deepgram Flux, which keeps its own turn detection.
- It is a per-call switch. You add turnDetection="true" to the ConversationRelay TwiML element.
- It reports turn state. With includeInPrompt="turnState", your application receives StartOfTurn, EagerEndOfTurn and EndOfTurn events over the WebSocket.
Twilio publishes no accuracy figure for the model. The claim is that interruptions drop, and the only way to check it is to listen to your own calls.
What does provider failover do when a speech provider fails?
Failover moves a live call to a backup provider so the caller stays connected. Twilio's changelog of November 18 says that if the speech-to-text or text-to-speech provider fails, "the call switches to its backup and stays connected," and your application keeps the same WebSocket connection.
The limits are specific, and they are the useful part:
| Behavior | What Twilio's changelog says |
|---|---|
| Speech-to-text and text-to-speech | Fail over independently |
| Switch count | Once for each per call, then the call ends if the backup fails |
| Speech during the switch | Recognition restarts and speech might be missed |
| Text not yet spoken | The backup voice speaks it, in order |
| Languages | Some have no default backup, so you must name one |
| Setup | A Failover element in TwiML, fixed for the call |

A backup provider keeps the line open. It does not guarantee that every word is heard during the switch, because Twilio says speech during the switch might be missed. For a collections call or an appointment confirmation, that gap is where an agent can ask the caller to repeat.
What else changed in the Conversation Relay release?
Four other items shipped in the same post. Dynamic Configuration Parameters let a team change speech hints, languages and voice settings in the middle of a call, so a caller who switches language does not force a restart. HD Audio gives the speech model clearer input. Advanced Conversational Voice adds ElevenLabs v4 Turbo, which Twilio says covers 90+ languages.
Twilio Agent Connect links external agent frameworks to Twilio voice and messaging. Twilio names AWS, Google Cloud, Microsoft Azure and OpenAI's speech-to-speech models as examples of stacks you can bring.
Conversation Intelligence is the supervision piece. Twilio's post describes a transcript-first view of AI conversations, with prebuilt operators in public beta that score task completion and detect hallucinations. Twilio labels those operators beta, so treat them as early.
Does your AI calling platform need these features?
Every AI calling platform needs an answer to the same three problems Twilio lists: callers cut off mid-thought, speech vendors that fail mid-call, and no view of what the agent said. A buyer should get an answer for each one, whether the platform is built on Twilio or not.
Ask four questions before you sign:
- How does the agent decide a caller has finished? A fixed silence timer and a semantic model behave differently on hesitant speakers.
- What happens when the speech provider fails mid-call? Ask whether the call survives and what the caller hears.
- Who owns the tuning? A developer platform hands you the settings. A managed service tunes them for you.
- What do you get after each call? A transcript lets you audit the agent yourself.
Topcalls is the managed route. A dedicated team configures the agent, so you need no prompt engineer, and a named account contact measures success against your KPIs. Every call returns a recording, a transcript, a summary, a disposition and a post-call analysis. AI calls start from $0.25 a minute, depending on the country you call, billed by the second, with the AI voice model, call recording, transcription and analytics included.
Our AI voice agents page lists what the agent does, and smart campaigns covers retries. For a use case, see AI appointment setting. Our build vs buy guide compares assembling a stack with a managed service, and our piece on AI voice latency explains why turn-taking feels slow. For the model side of the same stack, read the GPT-Live API guide.
To see what a managed campaign costs, run the ROI calculator or talk to us.
Sources
- Twilio blog, "From Voice AI Demo to Production", November 17, 2026.
- Twilio changelog, Smart Turn Detection general availability, November 13, 2026.
- Twilio changelog, speech-to-text and text-to-speech provider failover, November 18, 2026.
Frequently Asked Questions
Get AI calling tips in your inbox
No spam. One email per week with actionable sales automation tips.






