Best TTS for Voice Agents in 2026: Real Cost Per Minute, Not Per Million Characters

How much does TTS actually cost per minute of conversation?

Between about $0.009 and $0.18 per minute of agent speech, depending entirely on which voice you pick. Nobody in this category publishes that number, because they all publish dollars per million characters, which is a unit no one budgets a phone call in.

The conversion is trivial and the most useful thing in this post. Take the price per million characters, multiply by your characters per minute of synthesised speech, divide by a million. LiveKit's pricing calculator assumes 600 characters per minute of agent speech, and I use that throughout so my arithmetic reconciles with theirs. Rime's own page describes $0.03 per 1,000 characters as "~$0.03 per minute of audio," which implies roughly 1,000 characters per minute. Vendors don't even agree with each other on the constant.

That disagreement matters, because it's a 1.7x difference in every cost estimate depending on whose assumption you inherit. Measure your own: count the characters your agent actually synthesises per call from your logs, divide by agent speech minutes, and use that. It takes twenty minutes and makes every number below yours instead of mine.

Then apply the second correction, the one that saves people from over-budgeting. You pay for the agent's speech, not the call. On a four-minute call where the agent talks 40% of the time you're billing about 1.6 TTS-minutes, not 4. Estimates that assume the agent talks the whole call run roughly 2.5x too high, and I've seen that error survive all the way into a board deck.

Why is TTS the line item that decides your per-minute cost?

Because in LiveKit's own default calculator configuration, TTS at $0.0300 per minute is larger than STT, LLM, agent hosting and telephony combined. That's not my framing; it falls straight out of their published decomposition of a $0.0672 per-minute phone agent: agent session $0.0100, telephony $0.0100, LLM $0.0014, STT $0.0058, TTS $0.0300, observability $0.0100.

So the single highest-leverage cost decision in a voice stack is which voice you use, and it's usually made by whoever liked a demo. Moving from a $50-per-million-character voice to a $15-per-million-character one saves $0.021 per minute of agent speech. At 100,000 call-minutes a month with a 40% talk ratio, that's 40,000 TTS-minutes and about $840 a month, or roughly $10,000 a year, from one configuration line.

You can verify the arithmetic against LiveKit's own worked example on their pricing page. They price GPT-4o plus Nova-2 plus Eleven Multilingual v2 at about $0.2151 per minute, and $0.1800 of that, 84%, is the TTS, because Multilingual v2 lists at $300 per million characters. That example exists to show how the calculator works. It doubles as the clearest possible argument for reading this table before you pick a voice.

It's also the mechanical explanation for why a self-orchestrated stack lands near 2.5¢ a minute while LiveKit Cloud's own default lands at $0.0672. Three decisions do it: self-hosted agent workers remove the $0.0100 session and $0.0100 observability lines, third-party SIP replaces bundled telephony, and TTS selection replaces $0.0300 with $0.0090. Full working in the full LiveKit voice AI build guide.

Which TTS is fastest for voice agents?

Rime Mist v3, on published numbers, at 37ms time-to-first-audio p50 and 56ms p90. Rime is also the only vendor in this comparison publishing TTFA as a proper distribution with a stated measurement condition, and that condition matters enormously.

The condition is "at 1 concurrency." Rime states it plainly on their pricing page, an unusually honest disclosure that most aggregator posts strip out when they repeat the number. Your production concurrency isn't one. Published TTFA is a floor, not a forecast, and the gap widens exactly when you need the headroom.

Rime's other model, Coda, publishes 96ms p50 and 98ms p90: slower on the median but with a remarkably tight distribution, only 2ms between p50 and p90 against Mist v3's 19ms. For a voice agent, a tight tail is often worth more than a fast median, for the same reason p95 beats p50 everywhere else in the voice AI latency budget.

For self-hosting, Rime publishes sub-100ms model latency for Coda and roughly 70ms for Mist v3 on your own hardware, deployable via Docker Compose or Kubernetes. That's the only path in this category to removing network variance from your TTFA entirely, and at high volume it's where the next step-change in cost lives, though it converts a per-minute bill into a capacity-planning problem, a trade rather than a saving.

What actually breaks barge-in, and which providers handle it?

Mid-stream cancellation and word-level timestamps. A provider with better raw latency and no clean cancellation produces a worse-feeling agent than a slower one that stops when told, and this criterion only appears in comparisons written by people who've shipped an agent rather than benchmarked one.

Cancellation is the first requirement. When a caller interrupts, you must stop the in-flight synthesis, not merely stop playing the audio you already have. If you can't, you keep receiving audio for an utterance nobody wants, you keep paying for the characters, and depending on your buffering the caller may keep hearing it.

Word-level timestamps are the second, less obvious requirement. After an interruption you have to record what the caller actually heard, not what the model generated, or the agent refers back to information that was never delivered. Rime publishes this directly: Coda has word-level timestamps, Mist v3 doesn't. That makes it a barge-in correctness feature rather than a formatting nicety, and the reason I wouldn't choose a TTS on TTFA alone.

Both models stream over HTTP and WebSockets, which is table stakes. The thing to test before you commit is trivial and almost nobody does it: interrupt the agent at word five of a forty-word utterance and listen. The full mechanics, including the context-truncation code, are in turn detection and barge-in for voice agents.

from livekit.agents import AgentSession
from livekit.plugins import rime, cartesia, deepgram, elevenlabs

# Pick ONE. The rest of the session is identical in every case.
TTS = {
    # $0.0180/min @ 600 chars. 37ms TTFA p50. No word-level timestamps.
    "fast_cheap":  lambda: rime.TTS(model="mistv3"),

    # $0.0300/min. 96ms TTFA p50, 98ms p90 -- tight tail, timestamps: yes.
    "barge_in_safe": lambda: rime.TTS(model="coda"),

    # $0.0180/min @ 600 chars.
    "deepgram":    lambda: deepgram.TTS(model="aura-2"),

    # $0.0300/min.
    "cartesia":    lambda: cartesia.TTS(model="sonic-3"),

    # $0.0900/min -- 10x the cheapest option in this dict. Know why you chose it.
    "premium":     lambda: elevenlabs.TTS(model="eleven_flash_v2_5"),
}

session = AgentSession(
    tts=TTS["barge_in_safe"](),
    # ... stt, llm, vad, turn_handling unchanged
)

# Before committing: interrupt at word 5 of a 40-word utterance and listen.
# If audio continues, cancellation is broken and no benchmark number matters.

Which TTS should you actually pick?

Ranked by binding constraint, because "it depends" is not an answer and a single winner would be dishonest. One: if cost is binding, xAI Text to Speech at $15 per million characters is the cheapest credible option on a published rate card, at $0.0090 per minute of agent speech. Two: if latency is binding, Rime Mist v3 at 37ms TTFA p50. Three: if barge-in correctness is binding, and for a real conversational agent it should be, Rime Coda, for the word-level timestamps.

Four: if you need HIPAA or on-premise, Rime is the only provider in this comparison publishing both a self-host path (Docker Compose or Kubernetes) and an Enterprise BAA plus SOC 2 Type II report. Five: if voice quality is genuinely the product, a consumer brand, a character, something people choose to listen to, ElevenLabs is the quality benchmark and the 10x premium over the cheapest option is a real product decision, not a mistake. Say so in the budget rather than discovering it in the invoice.

Six: if you're multilingual, check the language list before the price. Rime lists eight production-ready languages on Coda and four on Mist v3, and a cheap voice that doesn't speak your customers' language isn't cheap. Seven: if you don't know yet, pick anything with a LiveKit plugin and make the swap a one-line change, the configuration in the snippet above.

The honest caveat: I haven't independently measured TTFA for these providers, and I won't publish numbers I didn't measure. Everything in the latency rows above is what the vendor publishes, labelled as such. Before you commit, run twenty synthesis requests of a fixed sentence against your two finalists from your own region at your own concurrency and compare. That measurement beats every published table, including this one.

To size the decision in money rather than cents, put your own volume and talk ratio into the voice AI cost calculator; TTS is the input that moves the output most. And if you'd rather have the whole pipeline built with the provider abstraction and cost instrumentation already in place, that's what voice AI development is. Under about 20,000 minutes a month, though, stay on a managed platform and spend the attention elsewhere; the reasoning is in the voice AI build vs buy break-even.