Rime Prism vs Inworld Realtime TTS 2: 200 Locales Against One Checkpoint

Here is a fact that should end most versions of this comparison before anyone opens a benchmark: Inworld Realtime TTS 2 publishes coverage in more than 200 locales and ships 282 built-in voices, while Rime Prism publishes no language count and no voice count at all. If you are deploying a voice agent into a dozen markets, that is not a trade-off to weigh — it is a filter that removes one option. The comparison only stays interesting for the buyer who is single-market and regulated, because Rime Prism is the only model in this pair that lets an application read and block a sentence before the caller hears it. Inworld Realtime TTS 2 is a text-to-speech model with a very good realtime story attached, and it has no equivalent checkpoint, because by the time it sees text the decision has already been made.

The names are also misleading in a way that has already produced bad advice. "Inworld Realtime TTS 2" is not a speech-to-speech model. Rime Prism is. That distinction sits underneath every number below and is worth saying before the numbers start.

One of these is not an agent

Inworld Realtime TTS 2 is text in, audio out. Duplex behaviour — hearing the caller while speaking, handling interruptions — comes from Inworld's separate Realtime API, which speaks the OpenAI Realtime protocol and advertises roughly 100 ms interrupt latency. That is a second product with a second integration, not a property of the TTS model itself. Anyone who reads "Realtime" in the model name and assumes end-to-end speech-to-speech has misread the catalogue.

Prism is the other shape entirely. Audio in, audio out, in one system, with a text waypoint in the middle and tool calls that run alongside speech rather than before it. It is priced per minute of call audio for exactly that reason — you are buying the call, not the text.

So the honest framing of this matchup is: Inworld Realtime TTS 2 is the best-in-class synthesis hop for a cascade you assemble, and Prism is a replacement for the cascade. You can prefer Inworld and still build a Prism-shaped agent, just in more pieces. You cannot prefer Prism and get Inworld's language coverage, because it is not published.

The coverage numbers, and what they are worth

Inworld's list is unusually specific, which is itself informative:

• Locales — 200-plus, split into 15 at Tier 1 quality and 79 at Tier 2.

• Voices — 282 built-in, with 20 curated presets.

• Cloning — zero-shot from 5 to 15 seconds of audio; Professional Voice Cloning in preview and flagship-only.

• Continuity — a single consistent voice identity carried across languages, which is the feature that makes a multi-market brand launch look deliberate rather than translated.

• Audio — MP3, WAV, OGG, FLAC and PCM output, 8 kHz to 48 kHz, 24 kHz default.

• Throughput — 2,000 characters per request, WebSocket streaming with 20 concurrent connections and 5 contexts.

• Latency — server-side P99 time to first byte under 100 ms for TTS-2; the Flash variant launched August 31, 2026 claims under 25 ms, with third-party measurement near 25 ms.

The Tier 1 and Tier 2 split is the part that deserves attention rather than the 200-plus headline. Fifteen locales at flagship quality out of 200 supported is a genuinely honest disclosure, and it tells you where the product is actually finished. If your markets are English, Spanish, Portuguese, German, French and Japanese, you are likely in Tier 1 and the headline is real. If they are the other 185, you are in the part of the catalogue that exists but has not been curated.

Compare that to Prism, where the absence of a published list is not evidence of poor coverage — Rime publishes language and voice counts for its separate text-to-speech products but not for Prism — yet it is evidence that you cannot plan against it. For a single-market deployment in the language Rime builds for, that absence costs you nothing. For anything broader, it is the whole answer.

The duplex question nobody asks

"Realtime TTS" and "speech-to-speech" sound like the same promise and behave nothing alike under interruption.

In a cascade built on Inworld Realtime TTS 2, interruption handling is a coordination problem across three services. The recogniser must notice the caller started talking, the orchestrator must cancel the language-model turn, and the synthesiser must stop mid-stream and clear its buffer. Inworld's Realtime API does the synthesiser's half well — roughly 100 ms to interrupt — but the other two halves belong to you, and the failure is audible: the agent keeps talking for a beat after the caller has begun, then restarts a sentence from the wrong place.

Prism handles this inside one model, and Rime names interruption handling as a design partner talking point. That is not a benchmark you can read; it is a class of bug you stop having. For a support line where callers talk over the agent constantly, it is worth more than a few points of Elo.

The counter-argument is that a well-built cascade with an aggressive cancel path is genuinely good at this, and the coordination code is a known quantity that many teams have already written. Interruption handling is hard, not impossible, and the difference between the two architectures narrows with engineering effort. What does not narrow is the number of services you are paged for at 3 a.m.

Price: four cents versus four cents, sort of

Inworld prices per character, Prism prices per minute of call audio, and the conversion between them is where honest comparisons go wrong.

• Inworld Realtime TTS 2 — $25 per million characters on-demand, $12.50 on Growth, down to about $5 at enterprise volume. TTS-2 Flash is $15, $7 and under $5 on the same tiers. At roughly 1,000 characters per minute of speech, the on-demand rate is about $0.025 per minute of generated audio.

• Rime Prism — $0.080 per minute of call audio on Flex, $0.064 on Scale, $50 of free credit on the free tier, custom on Enterprise.

So a minute of agent speech costs about 2.5 cents from Inworld and 8 cents from Prism. That is the number the comparison tables print, and it is misleading for two reasons.

The first is that Inworld's cents buy one hop. A cascade adds recognition — typically half a cent to a cent a minute — and a language-model turn, which at conversational volumes lands in the low single digits. Assemble the three and you are near six to nine cents per agent-spoken minute, which brackets Prism's eight. The gap in the table is an artefact of comparing a component to a system.

The second is that Prism's minute is not a minute of agent speech. It bills call audio, so a caller who talks for four minutes and hears the agent for one costs Prism five minutes worth. Cascade pricing scales with how much the agent says; Prism's scales with how long the call runs. On a support queue with long customer turns, that is a real premium. On a booking flow where the agent does most of the talking, it is close to a wash.

Inworld's plan structure adds a practical constraint worth checking: a Creator plan at $25 a month, $25 per million characters on demand, $12.50 on Growth, enterprise custom. Concurrency is 20 WebSocket connections on the streaming path, which for a mid-size queue is adequate and for a large one is the binding limit.

Control: one has a place to stand, one does not

This is where Prism earns its premium, and it is a structural difference rather than a feature gap.

Inworld Realtime TTS 2 is genuinely sophisticated about context. It reads prior turns as audio input, it takes free-form voice direction in plain English, it supports inline tags and a per-character spelling mode for order numbers, codes and addresses. That spelling mode is the closest thing in this comparison to a pre-speech control, and it is good — a synthesiser that can say "four-seven-two" instead of "four hundred and seventy-two" solves a real class of support-call failure.

But it acts on text you have already committed to. There is no point in the turn where Inworld tells your application "here is what I am about to say" and waits. The spelling mode fixes pronunciation, not content.

Prism tells you, every turn. Rime's claim is that the text is a checkpoint rather than an intermediate: a guardrail can inspect every word before it is spoken, a policy layer can rewrite a rejected sentence and the model speaks the replacement, a prohibited claim can be blocked outright while the call continues. Whether the model's reasoning is genuinely audio-native is a claim worth testing; whether the text arrives before the audio is a design fact.

For a regulated single-market deployment, that fact is the reason to pay eight cents instead of assembling a cheaper cascade. For everything else, it is a feature you are buying and may not need.

What the leaderboard claims actually say

Inworld markets Realtime TTS 2 as number one on the Artificial Analysis Speech Arena. Independent arena listings put it at an Elo of roughly 1,245, ranked third on the current board, with TTS-2 Flash at around 1,210 and sixth. A top-three finish on a public arena is an excellent result and a more useful claim than the one in the marketing; the discrepancy is a positioning statement colliding with a live table that moves as voters arrive.

There is no equivalent for Prism. Its 56.6 on τ-Voice is an agentic task-completion score that Rime ran itself, on a harness Artificial Analysis also uses, with no published leaderboard entry at launch. The two numbers measure different things — how a voice sounds to human raters versus how often an agent completes a simulated service task — and there is no arithmetic that turns one into the other.

That asymmetry is, again, the comparison in miniature. Inworld has a voice quality score with a referee. Prism has an agent capability score without one. Neither company has published a head-to-head, and neither model has ever been run on the same task by the same lab.

If you build the cascade, we route the middle

Inworld Realtime TTS 2 is sold direct by Inworld and through several third-party platforms. Rime Prism is sold direct by Rime. Neither is on OrcaRouter's catalogue today, and this piece does not claim otherwise.

The cascade architecture that Inworld's model belongs to still needs a language model in the middle, and that middle is where quality and cost are actually decided. OrcaRouter puts 200-plus models behind one key at 0% markup over provider list price, with automatic failover when a provider degrades and a routing DSL for sending different turns to different models — cheap for the acknowledgement, capable for the lookup, all through one endpoint. For a voice agent that is the useful unit: one place to change the model without touching the call path.

We also route synthesis, if you want it on the same key rather than a second contract: openai/tts-1-hd and google/gemini-3.1-flash-tts-preview are both available at list price. Neither is Inworld Realtime TTS 2. Both are the last hop of a cascade you can assemble today.

Two questions, and the one that decides it

Question one: how many languages? If the answer is more than a handful, Inworld Realtime TTS 2, and the comparison ends. Two hundred locales with a published Tier 1 list of fifteen beats an unpublished list with no numbers attached, and you cannot run a procurement against a blank field.

Question two: does anyone need to approve the sentence before it is spoken? If yes, Prism. Not because it is better at speech — there is no evidence either way — but because it is the only model here that produces the sentence as an artifact your application can hold. That is the whole product, and it costs about eight cents a call minute.

If neither binds, the honest answer is that you should build the cheaper cascade around Inworld's synthesiser and spend the savings on interruption handling, because that is the part of the experience a caller actually notices. Two and a half cents a minute for the synthesis hop is a strong offer from a model ranked second on a public arena, and the checkpoint you are declining is one you have just told us you do not need.

A note on the figures before anyone quotes them: Inworld's coverage, latency and price numbers are Inworld's own, the arena Elo is Artificial Analysis', and it disagrees with Inworld's number-one positioning. Prism's 56.6 is Rime's, unreproduced. No independent party has measured either model doing the other's job, which is the one gap neither company has closed.