Rime
by Rime Labs
Text-to-speech trained on studio-recorded conversation, built for regulated call volume
Rime builds conversational text-to-speech models for enterprise voice agents in healthcare, financial services, food service and telecom. Rather than scraping web audio, it records its own conversational data in a San Francisco studio, which is what lets it hold pronunciation of brand names and industry terms deterministically. It ships as a cloud API, a dedicated VPC, or fully on-premises containers for teams whose audio cannot leave their network.
Rime is an enterprise text-to-speech vendor built around a contrarian data decision: instead of scraping web audio, it runs a recording studio in San Francisco to collect its own conversational speech, on the argument that conversational responses have shorter breath groups and different stress patterns than the read-aloud speech most TTS corpora contain. Founded in 2022 by Lily Clifford, Brooke Larson and Ares Geovanos — a Stanford PhD student, an ex-Amazon Alexa engineer and a Stanford engineer — the roughly 35-person company now ships three production models. Coda is the flagship default, with the highest voice-quality scores, 253 voices, word-level timestamps and coverage of nine languages including Arabic, Hindi, Japanese and Portuguese. Mist v3 is the latency model, quoting time-to-first-audio around 37ms at P50 across 78 voices in English, French, German and Spanish. Mist v2 remains available because it, unlike v3, supports inline pronunciation control and custom pauses. That deterministic control over proper nouns and brand names is the commercial pitch: in regulated call flows, saying a drug or a policy name wrong is a compliance event, not a cosmetic flaw. Deployment is unusually flexible for the category — cloud API with regional endpoints, a dedicated VPC, or fully on-premises via standard container images under Docker Compose or Kubernetes, with audio and text never leaving the customer network. Rime is SOC 2 Type 2 compliant as of May 2025 and HIPAA compliant since February 2024, and never trains on customer data. It raised a $24M Series A led by M13 in July 2026 after a $5.5M seed, and handles over 100 million calls monthly for Mayo Clinic, Dialpad, Upstart and Asurion.
A voice-agent or contact-centre engineering team in healthcare, insurance or fintech that is blocked on pronunciation accuracy or on getting audio out of a third-party cloud — Rime is one of the few TTS vendors that will run fully inside your own network under a BAA.
Deterministic control over how brand names, drug names and account terms are pronounced, at sub-100ms time-to-first-byte, without shipping customer audio to a vendor cloud.
At a Glance
- Category
- Audio & Voice
- Pricing
- Usage-based, Freemium, Contact for pricing
- Target Market
- CTOs, Enterprise Developers, ML Engineers, Contact Centre Technology Leaders, Healthcare IT
- Deployment
- Cloud-first, Self-hosted, API-based, Hybrid
- Founded
- 2022
- Headquarters
- San Francisco, United States
- Team Size
- 11-50
- Customers
- Over 100 million calls handled monthly; named customers include Mayo Clinic, Dialpad, Upstart, Asurion, ConverseNow, CallRail, Telnyx and NLX
Key Features
- ✓Studio-recorded proprietary training data
Rime records its own conversational speech rather than scraping the web, which is why breath groups, stress patterns and currency read correctly.
- ✓Deterministic pronunciation control
Proper nouns and brand names can be pinned rather than left to the model, turning a compliance risk in regulated calls into a config entry.
- ✓Coda flagship model
253 voices across nine languages with the highest voice-quality scores, word-level timestamps and a spell() function for character-by-character readout.
- ✓Mist v3 low-latency model
Roughly 37ms P50 time-to-first-audio and time-to-first-byte well under 100ms across 78 voices, for agents where turn-taking speed dominates.
- ✓On-premises and VPC deployment
Standard container images under Docker Compose or Kubernetes keep audio and text inside your network end to end — rare among TTS vendors.
- ✓Compliance posture
SOC 2 Type 2 since May 2025 and HIPAA compliant since February 2024, with no customer text or audio used for model training.
- ✓Voice-framework integrations
Drops into LiveKit, Pipecat, Vapi, Daily, SignalWire and VideoSDK, so it can replace an incumbent TTS without rebuilding the agent.
Capabilities
Use Cases
- •Healthcare patient calls
Mayo Clinic uses Rime where mispronouncing a drug or clinician name is a clinical and compliance problem, not a cosmetic one.
- •Contact-centre voice agents
Dialpad and Asurion run customer conversations where sub-100ms time-to-first-byte keeps turn-taking from feeling like an IVR.
- •Lending and fintech verification calls
Upstart uses Rime for flows where account numbers, currency amounts and legal terms must be read back exactly right.
- •Regulated deployments with data-residency constraints
Banks and health systems run Rime containers inside their own VPC or datacentre so no call audio reaches a vendor cloud.
- •Outbound survey and research calling
AAPOR 2026 research across nearly 100,000 automated calls found Rime voices produced the highest caller retention.
Ideal For
Best For
- ✓Healthcare and insurance voice agents that need HIPAA coverage plus a BAA and cannot tolerate mispronounced drug or policy names
- ✓Teams running fully on-premises or VPC deployments where call audio and text must never leave their own network
- ✓Real-time conversational agents where time-to-first-byte, not broadcast polish, is the binding constraint
- ✓Contact centres with heavy proper-noun load — brand names, street addresses, currency amounts — that generic TTS gets wrong
- ✓Voice-framework builds on LiveKit, Pipecat, Vapi, Daily, SignalWire or VideoSDK that need a drop-in TTS provider
Not Ideal For
- ✗Audiobook, dubbing or broadcast narration work — Rime optimises for conversational prosody and speed, where ElevenLabs is the stronger pick for studio-grade polish
- ✗Products needing broad multilingual reach: Coda covers nine languages and Mist v3 only four, well short of the 30+ that ElevenLabs and the hyperscalers offer
- ✗Teams that expect a maintained official SDK — Rime's own docs state it publishes no npm or PyPI package, so you write raw HTTPS calls
- ✗Buyers who shortlist on third-party review evidence: Rime has effectively no G2, Capterra or TrustRadius footprint to check against
Deployment
Market & Ratings
Over 100 million calls handled monthly; named customers include Mayo Clinic, Dialpad, Upstart, Asurion, ConverseNow, CallRail, Telnyx and NLX
Market Analysis
Pros
- ✓Pronunciation accuracy is a genuine architectural advantage, not marketing — the studio-recorded corpus is why currency, addresses and proper nouns read correctly
- ✓One of very few TTS vendors offering fully on-premises and dedicated-VPC deployment, which unblocks health and financial buyers that cloud-only rivals cannot serve
- ✓Strong published latency: ~37ms P50 time-to-first-audio on Mist v3 and sub-100ms model latency on Coda
- ✓Transparent per-character pricing with a substantial free allowance and no credit card required
- ✓Credible enterprise proof — Mayo Clinic, Dialpad, Upstart and Asurion, plus 100M+ calls a month and an AAPOR study across ~100,000 calls showing the highest caller retention
Cons
- ✗No official SDK: Rime's own documentation states it publishes neither an npm nor a PyPI package, so integrations are raw HTTPS against a single endpoint
- ✗Language coverage is narrow — nine languages on Coda and only English, French, German and Spanish on Mist v3, against 30+ from ElevenLabs and the hyperscalers
- ✗Capability is split across models, so teams must trade off: inline pronunciation control exists only on Mist v2 and v1, not on the fastest model
- ✗Almost no independent review footprint — no readable G2, Capterra or TrustRadius profile, and the Hacker News presence is largely founder-authored posts, so third-party validation is thin
- ✗Model naming has churned (an 'Arcana' model promoted in blog posts does not appear in current API documentation), which creates migration risk for teams pinning a model ID
- ✗Not the choice for broadcast-quality narration — the design point is conversational speed, where reviewers still rate ElevenLabs higher on polish
Pricing
Starter
From $0.03 / 1K characters
- ✓Mist v3 at $0.03 / 1K characters (~$0.03/minute)
- ✓Coda at $0.05 / 1K characters (~$0.05/minute)
- ✓~800 minutes free (about 800k characters) plus 3,000 free minutes on the Starter plan
- ✓20 concurrent TTS generations
- ✓Public Slack support
- ✓No credit card required to start
Enterprise
Contact for pricing
- ✓Custom volume pricing
- ✓Unlimited concurrent TTS generations
- ✓Unlimited custom TTS voice clones
- ✓Cloud, on-prem or VPC deployment
- ✓BAA (HIPAA) and SOC 2 reports
- ✓SLAs and dedicated support
Rime publishes real list pricing, which is unusual in enterprise voice: metering is per character, at $0.03 per 1K characters for Mist v3 and $0.05 for Coda — roughly a cent per minute either way — so the model you pick is a direct cost lever. The free allowance is about 800 minutes plus 3,000 Starter minutes with no credit card, enough to load-test before committing. Starter caps you at 20 concurrent generations, which is the practical ceiling that pushes production call centres to Enterprise; Enterprise removes concurrency and voice-clone limits and is where on-prem or VPC deployment, SLAs, the HIPAA BAA and SOC 2 reports live, all at custom volume pricing.
Security & Compliance
Connect
Sources
This page was written from 8 sources, 5 on domains other than rime.ai.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Retell AI
Developer platform for building, deploying and monitoring AI phone agents, billed per minute
AssemblyAI
Speech-to-text, speech understanding and voice agent APIs priced per hour of audio
Cresta
Contact center AI platform for autonomous AI agents, real-time agent assist and conversation intelligence
Deepgram
Voice AI API platform for real-time speech-to-text, text-to-speech and voice agents