Twilio alternative
A Twilio alternative built for AI voice agents.
Most teams looking for a Twilio alternative are not unhappy with Twilio’s phone network. They are tired of the glue around it: a WebSocket server for call audio, separate speech-to-text, LLM and text-to-speech vendors, turn detection, interruptions and transfers, all written and run by hand.
VocalStack Voice is the layer that replaces the glue. SIP, WebRTC and PSTN under one API, speech and model providers wired by configuration, and barge-in, DTMF and handoff to a person built in. Your Twilio numbers can stay: Twilio is one of the carriers it connects to.
Median 247 ms, P95 under 300 ms, phone to phone · SIP · WebRTC · PSTN · India/EU routing · Real-time PII redaction
- AI agent00:00
Hi, this is Larkspur Home about your sofa delivery. We have you down for tomorrow between nine and
- Caller00:04
Sorry, can we make it Saturday? I am out tomorrow.
- System00:05
Barge-in · agent stopped speaking and replanned
- AI agent00:06
Saturday works. I have nine to one, or two to six.
- Caller00:09
Two to six. And the floor model I saw had a scratched armrest. Is there a discount on that one?
- System00:14
Handoff to customer care · order, new slot and transcript attached
- Human00:22
Hi, Leah from customer care. I have your Saturday slot here, and your question about the armrest.
What “just build it on Twilio” usually means.
As of Oct 2026, Twilio’s bidirectional Media Streams send a call’s raw audio to a WebSocket server you run, and play back whatever audio you send. Everything between those two points is yours to build, connect and keep running.
- 01Twilio numberTwiML opens a bidirectional stream.
- 02Your WebSocket serverReceives raw audio, sends audio back. You host and scale it.
- 03Speech-to-text vendorA second contract, a second connection.
- 04Your LLM and turn logicWhen to speak, when to stop, when to transfer.
- 05Text-to-speech vendorA third contract, streamed back in time.
- 06Back to the callerEvery hop adds to the delay they hear.
- Every hop is a vendor
- Separate keys, quotas, outages and invoices for the network, speech recognition, the model and the voice.
- Every hop adds delay
- The caller hears the sum of all of them, and you only see it if you instrument each one.
- The hard parts are yours
- Interruptions, keypad input, mid-call context and transfers to a person are code you write and maintain.
Latency, measured phone to phone.
Conversation runs on short gaps. Across ten languages, the average gap between one speaker finishing and the next starting is about 200 milliseconds. VocalStack Voice publishes its own end-to-end figures against that bar.
- Human conversation, mean gap between turns
- 208 ms
- VocalStack Voice, median
- 247 ms
- VocalStack Voice, P95
- < 300 ms
VocalStack Voice figures are VocalLabs’ own, as published on the VocalStack Voice product page: end to end, phone to phone, measured at five capture points from the phone edge through ASR, LLM and TTS. Human figure: mean response offset across 10 languages, Stivers et al., PNAS 2009. · Source: Stivers et al., PNAS 2009
Twilio alternative or Twilio add-on: DIY vs VocalStack Voice.
| DIY on Twilio | VocalStack Voice | |
|---|---|---|
| Call audio to your AI | Media Streams to a WebSocket server you host | Routed over edge media servers to the models you choose |
| Speech, model and voice | Separate vendors you contract and connect | Wired by configuration: OpenAI Realtime, Deepgram, ElevenLabs, Sarvam, Azure, Google and others, or your own models |
| Interruptions | Turn detection and playback cancelling you write | Barge-in built in |
| Keypad input | DTMF arrives as events your server handles | DTMF built in |
| Handoff to a person | Transfer logic and context passing you build | Escalation with full conversation state and transcript |
| Latency | The sum of your chain, measured by you | Median 247 ms, P95 under 300 ms, live in the dashboard |
| Data controls | Set up separately with each vendor in the chain | India/EU routing, real-time PII redaction, consent tones, encryption |
| Carriers | Twilio | Twilio, Exotel, Ozonetel, Airtel IQ, Plivo, Kaleyra or SIP trunks |
Twilio column describes Programmable Voice with bidirectional Media Streams as documented on twilio.com, as of Oct 2026. Twilio also offers ConversationRelay, which manages speech recognition and voices for you; see the questions below. · Source: Twilio Media Streams docs
What you stop building.
- 01Turn-taking
Barge-in
Callers can cut in mid-sentence. The agent stops talking and responds to what they said.
- 02Input
DTMF
Keypad input alongside speech, for PINs, account numbers and callers who prefer to press a key.
- 03HITL
Human handoff
Escalate to a person at any point with the full conversation state and transcript.
- 04State
Live context updates
Update the agent’s memory and state mid-call to keep it aligned with your systems and rules.
- 05Audio
Noise cancellation
Calls are denoised in real time, with no extra setup.
- 06Observe
Events and metrics
Webhooks for call events and transcripts, and metrics streamed to your data stack.
Keep Twilio. Change what sits on top of it.
VocalStack Voice is carrier-agnostic. Twilio is on the carrier list next to the rest, so moving the agent does not mean moving your numbers.
Moving a voice agent off a DIY Twilio stack.
- 01
Connect your carrier
Add Twilio, or whichever carrier or SIP trunk you use, as the carrier for VocalStack Voice. Existing numbers keep working.
- 02
Choose the models
Keep the speech, LLM and voice providers you use today, switch some of them, or host your own next to the telephony.
- 03
Move the logic
Point your prompts, tools and transfer rules at VocalStack Voice through its REST API and webhooks. The WebSocket glue between vendors is no longer yours to run.
- 04
Test under load
Simulate concurrent calls, check latency and verify the agent’s answers before any customer hears it.
- 05
Shift traffic gradually
Send a share of live calls through VocalStack Voice, compare against your current stack, then move the rest.
Questions buyers ask first.
Keep the carrier. Drop the glue.
Tell us how your agent runs today: carrier, models, call volume and where calls go when the AI is stuck. We will connect it to VocalStack Voice on a test number, so you can compare latency and handoff on your own calls.
