Speech to speech voice AI

Speech to speech voice AI: one model that hears your caller and talks back.

Speech to speech voice AI is a voice agent where a single AI model listens to the caller's audio and answers in audio, with no separate transcription or text to speech step. On a phone line that means quick replies and natural interruptions. Leadlock agents reply in about a third of a second on Grok (median of 461 live turns, measured September 2026).

Book a 15-minute call with Dan See plans and per-minute rates

Speech to speech means two different things

Search for the phrase and most results are voice changers for video creators. That is one meaning. The other is the one this page is about: an AI agent that holds a live conversation. The table sorts out the four things people mix up.

Tool What goes in What comes out Live conversation? Used for
Voice changer A recording of someone speaking The same words in a different voice No. File in, file out. Voiceovers, video ads, dubbing
Voice cloning Samples of one person's voice A reusable copy of that voice No. It is a voice, not an agent. Brand voices, narration
Text to speech Written text Spoken audio of that text No. It reads, it does not listen. Audiobooks, the last step of a chained agent
Real-time voice agent A caller talking, live A spoken answer the model decided on Yes. It listens, decides and speaks. Answering phones, booking, follow-up calls

Leadlock builds the last one. It does not change or clone voices. If you need a voice changer, a creator tool is the right product. If you need the phone answered, read on.

How it works, and why it sounds different

Older voice agents chain three services: speech to text writes down what the caller said, a language model writes a reply, and text to speech reads it out. Every hand-off adds delay, and the caller's tone is lost the moment speech becomes text.

A speech to speech agent has one step: audio in, one multimodal model, audio out. The model hears the caller directly, so it can answer quickly and stop talking when the caller cuts in. For the full comparison, see speech to speech vs speech to text to speech.

Video: Speech-to-Speech Voice AI Explained: Natural Calls That Book, by Dan Gauerke, founder of Leadlock.

The speech to speech models, compared

Reply time is the median time for the agent to answer a turn, measured on Leadlock production calls from September 1 to 25, 2026. Prices are Leadlock's per-minute rates on the Starter, Agency and Pro plans and cover the voice model and voice.

Model Median reply time Per minute on Leadlock (Starter / Agency / Pro) Good to know
Grok Voice
xAI
313 ms (461 turns) $0.18 / $0.15 / $0.12 Fastest measured replies on Leadlock. A Premium tier (think-fast) runs at $0.21 / $0.18 / $0.15.
Gemini Live
Google
568 ms (187 turns) $0.18 / $0.15 / $0.12 Same entry rate as Grok Voice. A common pick for high call volume.
OpenAI Realtime
OpenAI
674 ms (1,568 turns) $0.25 / $0.22 / $0.20 gpt-realtime supports SIP, remote MCP and function calling. A Premium tier runs at $0.39 / $0.36 / $0.33.
GPT-Live
OpenAI
About 2.3 s (small sample) $0.39 / $0.36 / $0.33 Newer OpenAI voice model, API launched September 10, 2026, limited general availability. Too few calls yet for a firm number.
Hume EVI 3
Hume
Not measured by us Not offered on Leadlock A speech to speech model sold as an API, known for emotional range. You build the phone and booking layer yourself.

Two honest limits on this table. The time from call connect to the first greeting is longer than a reply turn: 0.7 to 1.5 seconds on every model we measured. And we publish no number for chained stacks, because we do not run one to measure. For a wider ranking that includes platforms, see the best speech to speech AI platforms, and for the two OpenAI models side by side, GPT-Live vs GPT-Realtime.

What it costs

A Leadlock bill is the plan fee plus minutes at that plan's rate. For 1,000 minutes a month on Gemini Live or Grok Voice on Starter: $97 + 1,000 x $0.18 = $277 a month. For 10,000 minutes on Pro: $497 + 10,000 x $0.12 = $1,697 a month. There are no separate speech to text, language model or text to speech fees.

Phone numbers and carrier minutes are billed by your own Twilio account, at about $0.013 a minute inbound, $0.018 a minute outbound and $1.15 a month per local number at Twilio's US rates. See every plan and rate.

Put one on a phone line in about 5 minutes

  1. Start the trial

    7 days and 50 voice minutes on any plan. A card is required.

  2. Create an agent

    Write a short prompt and pick the model: Gemini Live, Grok Voice or OpenAI Realtime. How to create an agent.

  3. Connect a number

    Connect your Twilio account, import a number and assign it to the agent. No SIP trunk. How to import a number.

  4. Call it

    Ring the number, interrupt it, and tune from what you hear. Your first live call.

What businesses use it for

  1. Answering every call

    Inbound, overflow and after hours. The agent picks up, answers questions and takes the details.

  2. Booking on the call

    GoHighLevel works out of the box, and ServiceTitan or anything else with an API can be connected. See Leadlock for GoHighLevel and the ServiceTitan voice agent.

  3. Calling new leads back

    Outbound calls to a new form fill or a list, so a lead hears from you while they are still interested.

  4. Qualifying before a person steps in

    The agent asks the screening questions and passes on the callers worth your team's time.

Glossary

Speech to speech (S2S)
One model takes audio in and gives audio out. Also written speech-to-speech or voice to voice.
Chained, cascaded or pipeline agent
Three services in a row: speech to text, a language model, text to speech.
Speech to text (STT)
Software that writes down what was said. Also called transcription.
Text to speech (TTS)
Software that reads text aloud in a synthetic voice.
Multimodal model
A model that works with more than text. Here, one that hears and speaks audio natively.
Reply time
How long the agent takes to start answering after the caller stops talking.
Barge-in
The caller interrupting while the agent is speaking, and the agent stopping to listen.
Turn-taking
How the agent decides the caller has finished. See the turn-taking guide.
Voice changer
A tool that re-voices a recording. Sometimes also called speech to speech. Not what Leadlock does.

Sources

Frequently asked questions

What is speech to speech voice AI?

It is a voice agent built on one AI model that takes in the caller's audio and produces spoken audio back. There is no separate step that turns speech into text and no separate step that turns text back into a voice. Fewer steps means faster replies and a conversation that can be interrupted naturally.

Is speech to speech the same as a voice changer?

No. The phrase is used for two different things. A voice changer takes a recording of you and returns the same words in a different voice. A speech to speech voice agent holds a live conversation: it listens, thinks and answers on its own. Leadlock builds voice agents for phone calls. It is not a voice changer.

How is speech to speech different from text to speech?

Text to speech reads out text that something else wrote. It has no ears and makes no decisions. A speech to speech model hears the caller, including tone and pauses, decides what to say and says it, all inside one model.

How is it different from a chained voice agent?

A chained (also called cascaded or pipeline) agent uses three services in a row: speech to text, a language model, then text to speech. Each hand-off adds delay, and tone of voice is lost when speech becomes text. A speech to speech agent does the whole turn in one model.

How fast does a speech to speech voice agent reply?

On Leadlock, replies take about a third of a second on Grok Voice: a median of 313 milliseconds across 461 live turns, measured on production calls in September 2026. Gemini Live measured 568 milliseconds and OpenAI Realtime 674 milliseconds in the same period. These are reply times per turn, not time to the first greeting.

Which speech to speech models can I run on Leadlock?

Gemini Live from Google, Grok Voice from xAI, and OpenAI Realtime and GPT-Live from OpenAI. You choose the model per agent and can switch it later. ElevenLabs voices are also available as an option, but that path is not native speech to speech.

What does speech to speech voice AI cost on Leadlock?

Plans are $97, $297 and $497 a month (Starter, Agency, Pro). On top of the plan you pay a per-minute rate that covers the voice model and voice: Gemini Live and Grok Voice are $0.18, $0.15 or $0.12 a minute depending on plan, and OpenAI Realtime is $0.25, $0.22 or $0.20. Phone numbers and carrier minutes are billed by your own Twilio account.

Can the caller interrupt it?

Yes. Because the model is listening to audio the whole time, a caller can cut in mid-sentence and the agent stops and responds to what was said, the way a person would. How quickly it yields is a setting you can tune per agent.

Can I put a speech to speech agent on my business phone number?

Yes. You connect your own Twilio account, import a number on the Phone Numbers page and assign it to an agent. Incoming calls to that number then go to the agent. No SIP trunk is needed.

Can it book appointments during the call?

Yes. GoHighLevel works out of the box, so the agent can check a GoHighLevel calendar and book during the call. ServiceTitan and anything else with an API can be connected with a custom function.

Can it make outbound calls too?

Yes. The same agent can place calls, for example to follow up a new lead, as single calls or as a campaign over a list.

When is a chained pipeline still the better choice?

When you must run your own fine-tuned text model, when the brand depends on one specific cloned voice from a text to speech vendor, or when you need fine control over which model and voice handles each language. For most booking, intake and follow-up calls, speech to speech is the better fit.

Is Leadlock HIPAA compliant?

No. Leadlock is not HIPAA compliant and does not sign a business associate agreement. Keep patient health information off these calls whichever model you pick.

Can I try it before I pay?

Yes. Every plan starts with a 7-day free trial that includes 50 voice minutes. A card is required to start the trial.

Hear a speech to speech agent take a call

15 minutes with Dan. You will hear a live agent answer, handle an interruption and book, and we will work out which model fits your calls.

Book a 15-minute call with Dan