Compare speech-to-speech and text-to-speech voice agents on latency, control, cost and auditability, and learn which fits support calls and chat.
If you are adding a voice to your support agent, you will quickly meet two terms: text-to-speech (TTS) and speech-to-speech (S2S). They sound like small variations, but they are different architectures with different trade-offs in speed, control, cost and auditability. This article explains how each works, where each one fits, and how to decide without chasing hype.
Two architectures, one goal
Both approaches aim to let a customer talk and hear a helpful answer. The difference is what happens in the middle.
- TTS (the cascaded pipeline): the customer's speech is converted to text, a language model writes a text reply, and a TTS engine reads that reply aloud.
- Speech-to-speech: a single voice-native model takes audio in and produces audio out, without text as the required middle step.
Most teams already run the first pattern in some form. If your agent answers on a website chat or WhatsApp in text, adding TTS is a small step: the thinking stays the same and only the last mile changes. Speech-to-speech is a different design, closer to a live phone conversation than to a chatbot that talks.
How the TTS pipeline works
A TTS-based voice agent is a chain of separate stages:
- 01Caller speaks
- 02Speech is transcribed to text
- 03Language model writes a text reply
- 04TTS engine reads the reply aloud
Because text sits in the middle, you get a lot for free:
- A readable transcript of every turn, which is easy to search, rate, flag and review.
- The same brain as your text agent. Your knowledge base, persona, guardrails and integrations work as they do in chat.
- Model choice. You can pick the language model on its reasoning quality, and pick the voice separately on how it sounds.
- Swappable voices. Changing the TTS provider changes the sound without touching the logic.
The weaknesses are mostly about time and feel. Each stage adds delay, and the reply typically can't start until the text is ready. Tone, hesitation and emotion in the caller's voice are largely flattened by transcription, so the agent responds to words more than to how they were said.
How speech-to-speech works
A speech-to-speech model listens and speaks natively. Audio goes in, audio comes out, and the model handles turn-taking itself.
What this tends to give you:
- Lower perceived latency. There are fewer hand-offs between stages, so the conversation can feel more fluid.
- Natural interruptions. Callers can talk over the agent, and the model can stop and adapt, as people do.
- Better prosody. The model can respond to pacing, emphasis and tone instead of only to a transcript.
The trade-offs are just as real:
- Less modularity. Voice and reasoning come from one provider's model, so you generally cannot mix and match.
- Vendor dependence. If that service is unavailable, a text pipeline can often fail over to another model, whereas a voice-native session may have no equivalent fallback.
- Harder oversight. Even when a transcript is produced, the decision-making happens in audio, which makes it less natural to apply the same checks you use on text.
- Weaker fit with knowledge-grounded answers. Grounding replies in your own documents is straightforward in a text pipeline and needs more care in a voice-native one.
Side-by-side comparison
Neither column is simply better. The right question is which failure is more expensive for you: an awkward pause, or an answer you cannot trace.
What matters most for customer support
Support has different priorities from a general voice assistant. Weigh these five.
1. Accuracy and grounding
Support answers should come from your own policies, not the model's general memory. A cascaded pipeline lets you apply the same knowledge sources and guardrails you use in chat, so voice and text give consistent answers.
2. Auditability
When a customer disputes what was said, you want a clear record. Text-first pipelines make it easy to review, rate and correct conversations, then use the corrected answer to improve the agent.
3. Languages and accents
Check coverage for the languages your customers actually speak, not only the ones in the demo. Providers differ in how natural each language sounds, and some languages may fall back to a different voice than you expect. Test each one with real phrases, including names and product terms.
4. Cost
Voice usually costs more than text because it consumes more compute and runs for the length of the call. As a hypothetical, if a shop receives 300 chats a month and moves 100 of them to voice, those 100 conversations will likely dominate the bill. Estimate cost per minute, not per message, before you switch channels.
5. Channel reality
Many support conversations never need a voice at all. Website chat, WhatsApp and Messenger are text-first, and callers on those channels expect to type. Voice matters most on phone calls, in a voice widget, or where hands-free use is the point.
Choosing by use case
| Scenario | Leans towards |
|---|---|
| Phone support with FAQs and order lookups | TTS pipeline, for grounded answers and a clear transcript |
| Read-aloud replies on a website widget | TTS, since the text answer already exists |
| Conversational concierge or guided onboarding | Speech-to-speech, for natural pacing |
| Voice assistant that operates a dashboard hands-free | Speech-to-speech, where interruption handling matters |
| Regulated or high-stakes answers | TTS pipeline, for inspectability |
| Multilingual coverage with minority languages | Test both; coverage varies by provider |
You do not have to pick one for the whole business. Many teams use a cascaded pipeline for customer-facing support and a speech-to-speech experience for internal or assistant-style use.
A practical decision checklist
- List the five most common reasons customers contact you, and note which need accurate lookups.
- Decide whether you need a transcript of every call for review and training.
- Test your top languages with real names, order numbers and product terms.
- Run the same ten scripted calls through each option and compare delay, interruptions and wrong answers.
- Estimate cost per call minute at your expected volume.
- Decide what happens if the voice provider is down, and whether a text fallback is acceptable.
Doing this in AI SNS CHAT
AI SNS CHAT supports both patterns, and they live under Voice in the sidebar (Overview, Analytics, Voice agent, Avatar, Voice Playground, Inbound calls, Call history, Outbound calls and more).
The TTS route for your support agents
Your agents answer from their knowledge base and persona as usual, and the voice is a separate choice. The text-to-speech providers are Grok/xAI (the default), OpenAI (gpt-4o-mini-tts), ElevenLabs and browser speech. If you choose ElevenLabs, you can use its voice library, voice design and cloning. Because the text reply sits in the middle, the same guardrails and data sources apply, and conversations remain reviewable.
Use Voice Playground to listen before you commit. Try your real greeting, a policy answer and a product name in each language you serve. The voice stack covers 22 languages, and some of them use an English phone voice as a fallback, so test the ones that matter to you.
Phone calls
For Inbound calls, you connect a Twilio number to each agent. The setup includes a webhook self-test and auto-repair. Call history keeps the record. Outbound calls run as batch campaigns that need a Twilio number, require a test call first, and can be stopped mid-run. Saved SIP trunk settings exist, but a call path that uses them has not been confirmed, so plan around Twilio for now. WhatsApp voice and managed phone numbers are listed as coming soon.
Speech-to-speech
Synthia, the AI fleet manager available from the sidebar and as a bubble on dashboard pages, offers voice or text help. Its speech-to-speech mode uses the OpenAI Realtime API, with OpenAI only and no fallback provider. Grok voice-to-voice is also offered, but by default it answers only the account owner. That is the trade-off described above: a natural conversation, tied to one provider.
Avatar and cost
The Avatar page can turn a voice into a talking avatar, made with Simli lip-sync or from an uploaded clip. On billing, a text reply uses one credit, while credits per voice minute are set by the platform admin, so check your account and the pricing page before planning call volume. Failover between model vendors applies to text replies. Details on channels are on the connections page, and the full list of capabilities is under features.
Frequently asked questions
Is speech-to-speech always better than TTS?
No. It tends to feel more natural and interruptible, but it usually ties you to one provider and makes answers harder to inspect. For support that must follow policy and leave a clear record, a TTS pipeline is often the safer choice.
Can I use my existing text agent for voice?
With the TTS route, yes. The agent's knowledge sources and persona produce the text reply, and a TTS provider reads it aloud. Pick the voice in the Voice section and try it in Voice Playground first.
Does AI SNS CHAT offer a fallback for speech-to-speech?
Synthia's speech-to-speech uses the OpenAI Realtime API only, with no fallback. Text replies are different, because they can fail over across model providers according to the agent's settings and plan.
Do I need voice for WhatsApp or Messenger support?
Usually not. Customers on those channels expect text, and a well-grounded text agent is faster to review and cheaper to run. Add voice where the channel is a phone call or where hands-free use is the whole point.
How should I estimate the cost of voice?
Think in minutes, not messages. Text replies cost one credit each, but voice credits per minute are set by the platform admin. Estimate your expected call minutes and confirm the rate for your plan before launching.

