Rime is a text-to-speech service for conversational voice applications such as customer-service calls in banking, healthcare, and food ordering. Voice models are available through a cloud API or as a private deployment on your own infrastructure with Docker Compose or Kubernetes, with regional endpoints to reduce network latency.
Audio streams over HTTP or WebSockets from a single rime-tts endpoint. You choose a model and a speaker, then shape delivery with controls for speed, custom pauses, inline pronunciation, text normalization, and a spell() function for reading out letters and digits. The service plugs into voice-agent stacks through LiveKit and Pipecat integrations, a CLI, and an MCP server.
| Model | Languages | Voices | Notes |
|---|---|---|---|
| Coda | 9 | 253 | word-level timestamps, sub-100 ms model latency on the self-hosted GPU engine |
| Mist v3 | 4 | 78 | around 37 ms P50 time to first byte, custom pauses, no inline pronunciation control |
| Mist v2 | 4 | 138 | inline pronunciation control, 175 ms median latency on A10G GPUs |
Features
- Streaming output: audio streams over HTTP or WebSockets for real-time agents
- Voice library: more than 600 voices across industries and styles, with accent, pace, and tone controls
- Delivery controls: speed adjustment, custom pauses, inline pronunciation, text normalization, and spell()
- Word timestamps: the Coda model returns word-level timing for captioning and interruption handling
- Private deployment: runs inside your own environment with Docker Compose or Kubernetes when compliance requires it
- Agent integrations: LiveKit and Pipecat plugins, a CLI, and an MCP server for coding assistants
- Multilingual voices: dialects and accents across the supported languages of each model