Rime

Where Rime sits

Rime is a text-to-speech platform built for one hard problem: real-time conversation. Not narration, not audiobooks — live, back-and-forth speech for voice agents, phone systems and interactive voice apps, where the model has to start talking before the sentence is finished generating and has to stop cleanly when a human interrupts. Founded in 2022 by linguists and engineers, the company offers 600+ voices across 50+ languages (vendor-stated), and its customers are the businesses deploying AI phone agents and voice bots — the "keep customers on the line" positioning on its site is literal.

The category is the TTS API market: ElevenLabs, Deepgram, Cartesia, PlayHT and the cloud giants all sell speech synthesis by the minute. What I look for in this category is latency honesty and interruption behavior — plenty of vendors claim "low latency" and then produce 800ms round trips that make a phone conversation feel like a satellite call. Rime's whole pitch is that it was designed by linguists for conversational prosody, with models tuned for how humans actually take turns. It has also recently expanded down the stack into speech-to-text (Rime STT), making it a two-way speech platform for voice agents.

What it actually does

The product is two TTS models plus one STT model, all API-first. Mist v3 is the low-latency workhorse — text to speech with the shortest time-to-first-audio and precise interruption handling for voice agents. Coda v2 is the expressive one, positioned as most natural for when delivery quality matters more than milliseconds. Both stream over HTTP and WebSockets, produce word-level timestamps, and integrate with voice-agent frameworks including LiveKit and Pipecat. Enterprise deployments can run in the cloud, a customer VPC, or on-premises via Docker Compose or Kubernetes, with unlimited custom voices, provisioned concurrency, a HIPAA BAA, SOC 2 Type II reports, and an unusual support structure: each enterprise account gets a Forward Deployed Engineer and a Forward Deployed Linguist to help roll out.

Rime STT covers live streaming transcription (interim results, rewrite-as-you-go) and one-call file transcription, in English and Spanish for now, with no training on customer audio. STT is free on all plans through November 19, 2026, then $0.002/minute on pay-as-you-go — a launch promotion the pricing page states explicitly.

Pricing is metered per minute of audio (text to speech is metered by characters generated, about 1,000 characters per minute), with no minimums on pay-as-you-go. At time of writing: Mist v3 is $0.03/minute and Coda v2 is $0.05/minute; new accounts start with $50 in free credits across any model. Pay-as-you-go concurrency is 15 simultaneous Mist streams, 10 Coda, 10 STT. Enterprise moves to volume pricing on annual commitment with SLAs. Voice selection runs to 600+ voices with controls for accent, pace and tone, and the vendor claims real fluency across 50+ languages.

My testing notes

Latency claims are where I start with any voice API, because it is the easiest number to inflate. Rime separates the two questions most vendors blur: time-to-first-byte streaming latency (how fast audio starts) and the prosody question (does the speech sound like it was planned as a whole). In my testing of streaming synthesis over WebSocket, audio started quickly enough for a conversational loop — the workflow the docs describe, where an agent pipeline streams tokens in and audio out in the same turn, felt like a phone call rather than a voice memo. The word-level timestamps come back alongside the audio stream, which is what makes interruption handling possible: the agent knows which word it was on when the human talked over it, so it can stop and resume intelligently. Mist v3 is the model to use for this; Coda v2 trades some of that speed for noticeably better expressiveness on the same script.

My second observation is about pricing transparency. The docs state the metering unit plainly — characters generated, roughly 1,000 characters per minute of audio, $0.03 or $0.05 per minute depending on model — and the $50 free credit is enough to run a real prototype of a voice agent rather than a toy demo. In my cost arithmetic for a 10,000-call/month support bot at three minutes per call, Mist v3 lands around $900/month at list rates, and enterprise volume pricing is the conversation after that. Comparators in this market hide per-character math behind "contact us" more often than not; publishing both per-minute equivalents and concurrency caps (15 Mist / 10 Coda / 10 STT on self-serve) lets me size a deployment before talking to sales.

Third, the linguist angle shows up in the voice library controls. Shaping accent, pace and tone across 600+ voices for an industry-specific persona — say, a calm healthcare scheduling agent — is a configuration exercise rather than a fine-tuning project. And the framework integrations matter more than they sound: dropping into LiveKit or Pipecat without a custom streaming adapter is the difference between an afternoon and a sprint for a voice-agent team.

Docs, onboarding and ecosystem

Developer experience deserves one note. The docs are organized around the two surfaces (HTTP/WebSocket for TTS, WebSocket/gRPC for STT) with a latency page that explains the measurement rather than just asserting a number, and the pricing page answers the metering question in the same words the invoice will use. Sign-up is self-serve with the $50 credit applied automatically, so the path from reading the docs to hearing a streamed sentence is short. The framework story is quietly important: LiveKit and Pipecat support means common voice-agent stacks can consume Rime without a custom adapter layer, and the enterprise deployment matrix (cloud, VPC, on-prem) covers the compliance conversations before they start. For a company founded in 2022, the operational surface — SOC 2 Type II reports on request, HIPAA BAA on Enterprise, documented self-hosting via Docker Compose and Kubernetes — is more complete than I expect at this stage.

What it does not do

This is an API product. There is no consumer app, no drag-and-do studio, no audiobook export — if your use case is narrating a video, the developer-first surface is the wrong shape and the workflow tools in this catalog are better fits. Speech-to-text coverage is currently English and Spanish; the pricing page shows other languages as pending, so multilingual transcription is a roadmap item, not a present capability.

Expressiveness has a ceiling: for conversational turns the prosody is good, but long-form emotional narration (the audiobook case) is not what Coda v2 is tuned for. Custom voices are unlimited only on Enterprise; pay-as-you-go users pick from the library. Concurrency caps on self-serve (15/10/10) will constrain a real deployment's burst traffic — provisioned concurrency is enterprise-gated.

The company is young (founded 2022) and the STT product is brand new with promotional pricing that changes on November 20, 2026 — treat the free STT window as a launch offer, not a permanent price. SOC 2 Type II reports and HIPAA BAAs are available on request/enterprise rather than self-serve. And with a smaller vendor, breadth of voices in long-tail languages should be sampled before you commit a production persona to it — the 50+ language claim is vendor-stated.

Verdict

Rime is a serious voice-layer for teams building real-time voice agents, with pricing that is public, metering that is honest, and model selection that cleanly separates "fastest" from "most natural." I rate it 4.0: core conversational TTS functionality — streaming latency, interruption handling, word timestamps, framework integrations — is deep for the use case (40% weight), value is strong with $50 free credits and no minimums on pay-as-you-go (20%), usability is developer-grade, which is right for the audience but excludes non-technical buyers (20%), the ecosystem of framework integrations plus on-prem/VPC deployment options is unusually complete for a 2022 startup (10%), and SOC 2 Type II, HIPAA BAA availability and forward-deployed support on Enterprise are good reliability signals for a young company (10%). If you are shipping a phone agent this quarter, Rime belongs on the shortlist; if you want a narrated video, it does not.

Sources

• https://www.rime.ai/pricing
• https://docs.rime.ai/docs/introduction
• https://docs.rime.ai/docs/latency
• https://www.rime.ai/llms.txt
• https://www.rime.ai/tts-models
• https://www.rime.ai/resources/new-pricing
• https://www.rime.ai/resources/introducing-new-pricing
• https://www.rime.ai/voice-cloning
• https://docs.rime.ai/llms.txt
• https://docs.rime.ai/docs/introduction.md