Rime Prism vs Qwen-Audio-3.0-TTS: A Conversation Engine Against a Voice Layer

Rime Prism bills $0.080 a minute. A five-minute support call on it costs $0.40, and that $0.40 buys the listening, the reasoning, the turn-taking and the speech in one model. Qwen-Audio-3.0-TTS bills $19.30 per million characters, which for the same call is roughly $0.10 — for the speaking only. Put those numbers side by side and the comparison looks settled in Qwen's favour. It is not settled, because the two models are not competing for the same slot in a stack, and the $0.30 difference is mostly the price of deleting two handoffs from the path between a caller's sentence and an answer.

Rime launched Prism on 28 October 2026, the day before this page was written. It is a speech-to-speech model, not a text-to-speech model: it hears audio, works out a reply, writes that reply as text an application can inspect, and then speaks — while still listening. Alibaba announced Qwen-Audio-3.0-TTS at WAIC on 20 July 2026, in a Plus tier for quality and a Flash tier for real time. That model takes text and returns audio, full stop. Alibaba replaced it with Qwen-Audio-3.1-TTS on 23 September 2026, so the 3.0 in this headline is already the previous generation of its own family — a wrinkle that returns later and changes what you should actually buy.

The useful question is not which sounds better. It is which one you put in front of a caller, and what the rest of the diagram has to look like if you do.

Two products, one slot in the diagram

A conventional voice agent is a cascade: speech recognition turns the caller's audio into text, a language model decides what to say, text-to-speech turns that answer back into audio. Each hop is a place where the accent of the caller can be lost, where a filler word gets dropped, where 200 milliseconds get added.

Rime's argument is that the cascade throws away something it cannot recover. In a cascade the language model never hears how the caller said the thing — the sigh, the rising pitch at the end of a question, the pause before a credit-card number — because all it receives is a transcript. Prism keeps the acoustic signal intact all the way through the decision, then emits text as a checkpoint rather than as an intermediate format.

That is a real architectural difference and it is the whole basis of the product. It also means Prism is not a drop-in replacement for a TTS engine. Drop it into the last slot of a cascade and you have bought an expensive speech synthesiser that needs to hear the caller first. Drop Qwen-Audio-3.0-TTS into Prism's slot and you have a model that cannot hear anything at all.

What Rime Prism actually sells

Prism is the first model Rime has shipped that removes the STT and LLM stages rather than optimising one of them. The text layer is the feature enterprise buyers care about: a policy check, a PII filter or a human supervisor can read the exact words about to be spoken, and because the model is still listening while the check runs, the pause does not have to be audible. Tool calls run in parallel with speech rather than before it, which is a different way of hiding the latency of a database lookup than the usual "let me just check that for you" stall.

The published numbers, all of them from Rime: 0.57 seconds to first audio on the τ-Voice benchmark, 56.6% task completion on that benchmark, and 82.5% on its telecom subset. A Revmo trial on about ten hours of real ordering calls is cited as 100% order confirmations against 68% for an unnamed OpenAI model, with 85% of calls hallucination-free against 35%. Those trial figures come from a customer case study that Rime commissioned; the methodology is not published and nobody has reproduced it.

Pricing is per minute of call audio and is the same model at every tier: $0.080/min on Flex with no monthly fee, $0.064/min on Scale at $100 a month including $100 of credit that rolls over to $200, and volume rates on Enterprise. Free accounts get $50 of monthly credit. Prism streams over WebSockets and ships integrations for LiveKit and Pipecat. Self-serve SFT and RL fine-tuning APIs are in beta, not generally available; the Enterprise tier is where on-premises and VPC deployment, custom voices, a HIPAA BAA and SOC 2 Type II sit.

What Qwen-Audio-3.0-TTS actually sells

Qwen-Audio-3.0-TTS is a synthesiser with unusual control. It accepts inline delivery tags — [gasp], [laughs], [angry] — embedded in the text itself, so emotion is attached to a character rather than to a whole paragraph. It covers 16 languages and 20 Chinese dialects, renders up to three minutes in one request at 48 kHz, and comes with a library of more than 1,000 voices. The Plus tier is the quality tier; Flash targets real-time interaction and Alibaba's documentation puts its first-packet latency under 200 milliseconds, while earlier launch coverage quoted roughly 300.

The list price is $0.19253 per 10,000 characters in the Beijing region and $0.20 in Singapore, which is $19.30 to $20 per million characters — the figure Artificial Analysis carries on its leaderboard. Throughput is capped at 180 requests per minute in both regions. There are no weights, no self-hosting option and no on-premises story: if Alibaba's regions are a problem for your data-residency review, this model is simply not available to you.

The scoreboard

• Category — Rime Prism hears audio and speaks back; Qwen-Audio-3.0-TTS takes text and returns audio.

• Launch — Prism on 28 October 2026; Qwen-Audio-3.0-TTS on 20 July 2026, superseded by Qwen-Audio-3.1-TTS on 23 September 2026.

• Billing unit — Prism $0.080 per minute of call audio on Flex, $0.064 on Scale; Qwen $19.30 per million characters on Plus, about $13.75 to $15 per million on Flash.

• Latency — Prism 0.57 s time-to-first-audio, vendor-measured; Qwen Flash under 200 ms first packet per Alibaba's documentation.

• Independent score — Prism has none on Artificial Analysis's Speech-to-Speech Index; Qwen-Audio-3.0-TTS-Plus sits at Elo 1,269, sixth of 99 on the Text-to-Speech Arena as the board stood on 29 October 2026.

• Languages — Prism English only at launch; Qwen 16 languages plus 20 Chinese dialects.

• Deployment — Prism cloud, VPC or on-premises with HIPAA BAA and SOC 2 Type II on Enterprise; Qwen only through Alibaba Cloud Model Studio.

The five-minute call, priced

Assume a five-minute inbound support call in which the agent speaks about 1,000 characters a minute — Rime's own pricing FAQ uses that ratio to convert between its character-billed TTS and its minute-billed Prism, so it is a fair common denominator. That is 5,000 characters of speech.

On Prism the bill is five minutes at $0.080, or $0.40. At the Scale rate it is $0.32. Ten thousand such calls cost $4,000 and $3,200 respectively.

On a cascade built around Qwen-Audio-3.0-TTS-Plus, the 5,000 characters cost $0.0965. Add Rime's own speech recognition at $0.002 a minute — five cents for the call — and a text model for the reasoning. At a blended $0.50 per million tokens and a few thousand tokens per turn, that adds well under a cent. The cascade lands near $0.11 per call, or about $1,100 across ten thousand calls. Flash instead of Plus takes the synthesis line down to roughly $0.069 in Beijing, and the total to about $0.08.

So the cascade is roughly a quarter of Prism's price, and the gap is $2,900 to $3,200 per ten thousand calls. Those are list prices and the assumptions are ours, not Rime's; the point is the shape of the gap, not the third decimal.

The middle of that cascade is also the one component you can move without touching the other two. OrcaRouter routes text models behind a single API — 200-plus models, provider list price passed through at 0% markup, and automatic failover when a provider degrades mid-conversation — so the reasoning step can be re-pointed at a cheaper or stronger model between calls. Prism's text checkpoint is ordinary text for the same reason: whatever consumes it, from tool-calling to post-call summarisation to QA scoring, can sit on a routed endpoint and stay independent of the voice layer underneath.

Where the cascade still wins

Multilingual contact centres, for one. Prism is English-only at launch, with a limited voice set. Qwen-Audio-3.0-TTS ships 16 languages and 20 Chinese dialects — and its successor adds cross-language timbre transfer across Mandarin, Cantonese, English and Japanese. If a meaningful share of your callers are not speaking English, this comparison ends at the spec sheet.

One-shot generation is the second case. IVR prompts, outbound notifications, audiobook chapters, game dialogue: none of them involve a caller talking back, so there is nothing for a speech-to-speech model to hear. Paying $0.080 a minute for a model that can listen when nobody is speaking is paying for an unused half.

Volume is the third. At a quarter of the per-call cost, a cascade leaves budget for the parts that actually differentiate an agent — better retrieval, a stronger reasoning model, human escalation routing.

Where Prism still wins

Regulated inbound telephony is the case Prism was built for. A bank that must be able to show what its agent was about to say, and block it before a customer hears it, cannot do that with a fused realtime model and can only approximate it with a cascade whose text stage is a separate service with its own latency and its own failure mode. Prism puts the checkpoint inside the model and keeps listening through it.

It also wins on the metric that decides whether a call feels human: dead air. A cascade that is fast on paper still stacks endpointing, model time-to-first-token and synthesis time-to-first-packet, and a cascade that lands under a second end to end is doing well. Prism measures 0.57 seconds from the end of the caller's turn to the first audio of the reply, in one hop.

And it wins on deployment for the buyers who cannot use a Chinese cloud region at all — VPC, on-premises, a HIPAA BAA, SOC 2 Type II, and fine-tuning on your own call recordings rather than on a shared base.

The number neither side can give you

Neither model has the number a procurement team would most like to have: an independent head-to-head on the same benchmark.

Prism is not listed on Artificial Analysis's Speech-to-Speech Index. Its 56.6% τ-Voice figure is a vendor-run result, and the table Rime publishes alongside it is worth reading carefully: GPT-Live-1 (Sol, low) is plotted at 59.3%, higher than Prism, while the surrounding copy says Prism is "ahead of every realtime model from OpenAI and Google." Launch posts also cite "under 300 ms" and "4× faster than GPT-Live," figures that do not appear in the benchmark table, which says 0.57 s. Different measurements may explain the difference; the page does not.

Qwen's number is the stronger one, because 1,269 Elo comes from blind listener votes rather than from Alibaba. It is also no longer the best number in its own family: Qwen-Audio-3.1-TTS-Plus sits third at 1,305 as of 29 October 2026, and artificial-analysis revisions have moved this board more than once this year, so scores from different snapshots should never be subtracted from each other.

The caution specific to Rime cuts the other way. The only Rime voices that have been independently ranked are its TTS models, and on the same leaderboard Coda sits 65th at Elo 1,054 and Mist V2 92nd at 903 — the bottom half of 99. Prism may well sound better than its siblings; nothing published lets a buyer confirm that before signing.

What to watch

Three things would change the recommendation. Prism appearing on the Speech-to-Speech Index would give it the first independent number it has ever had. Full-duplex operation — Prism is half-duplex with barge-in at launch — would widen the gap against any cascade that has to wait for the caller to finish. And the self-serve fine-tuning APIs leaving beta is the difference between "enterprises can fine-tune" and "you can fine-tune on Tuesday."

On the other side, watch what Alibaba does with Qwen-Audio-3.1-TTS pricing after the roughly 70% cut it applied on 22 September 2026. A cascade whose middle and last hops are both getting cheaper is a moving target, and Prism's per-minute rate is fixed against it.

For most teams the honest answer is both, not either. Put Prism in front of inbound callers whose conversations need a checkpoint and a single-hop latency budget. Put Qwen-Audio-3.0-TTS — or, more realistically now, Qwen-Audio-3.1-TTS — in the outbound paths where nobody talks back and the bill is per character. The mistake is choosing one because a benchmark table said so, when the table has a footnote and the footnote is doing most of the work.