Rime AI

Rime's pitch is narrow on purpose, and that narrowness is the reason to consider it. Two models, two jobs. Mist v3 is the low-latency workhorse: 37 ms time-to-first-audio at P50 and 56 ms at P90, word-level timestamps, interruption handling and granular pause control. Coda is the expression model at 96 ms P50, and the pricing page is unusually candid about what it does not do yet: no word-level timestamps, and custom pause control listed as coming soon. That transparency matters more than a feature grid, because a voice-agent builder needs to know which model can support barge-in before committing. Pronunciation control is the second differentiator. Inline control renders names, addresses and alphanumerics the same way every time, so a bank never ships a demo where the model guesses at a customer's surname. Spell reads characters individually when a caller repeats a code, and continuous speed adjustment lets you tune pacing without re-recording. Both models support automatic text normalization and speed adjustment; the linguistic framing on the homepage (the company was founded by people who study language) is marketing, but the feature list backs it up. The deployment story is what large buyers should test first. Cloud, VPC and on-prem via Docker Compose or Kubernetes carry the same feature set, and self-hosted latency is quoted at sub-100 ms on Coda and around 70 ms on Mist v3 on your own hardware. HIPAA BAA and SOC 2 Type II reports are Enterprise-tier items, along with unlimited concurrency, unlimited custom voice clones, SLAs and a Forward Deployed Engineer plus a Forward Deployed Linguist. If you are a healthcare intake line or a cardholder hotline, those are the boxes that decide the deal. Where it falls short: the Starter tier caps you at 20 concurrent generations, which is fine for staging and tight for a busy production line. Production-ready languages differ by model (Coda lists English, Arabic, French, German, Hindi, Japanese, Portuguese and Spanish; Mist v3 lists English, French, German and Spanish), so the 50+ headline number needs checking against your target language and your chosen model. Coda lacks spell and word-level timestamps, so a use case that needs both expression and interruption handling has to split across two models or wait. And this is an API, not a voiceover studio: no drag-and-drop editor, no character-voice library for podcasters. Use it when latency and pronunciation accuracy are measurable business costs. Skip it if you are producing narration, or if you need one model that covers every call type without per-use-case tuning.