What’s new and what changed across the SignalWire platform.
To follow along in a feed reader, append .rss to this page’s URL.
Multilingual agents that switch languages and voices in real time
The new multilingual AI parameter accepts an array of language-voice pairs.
The AI agent continuously detects the other party’s language,
and switches to the configured TTS voice (and language) whenever it changes.
The multilingual and languages objects within the AI method are mutually exclusive.
See ai.multilingual.
AI agent text chat added
You can now chat by text with AI agents that were previously reachable only by voice. In the REST API, AI Chat is a single, unified JSON-RPC-style Send AI chat methods endpoint with six methods.
The Python Server SDK wraps the endpoint in AIChatClient
and ChatGateway.
JSON-RPC errors are listed with the rest of the error codes.
Default TTS engine changed from Google to ElevenLabs
Voice id strings
that don’t specify an engine now resolves against elevenlabs instead of gcloud.
Always declare your engine
Default engines can change. Ensure your agent continues using its configured voice by always declaring the engine explicitly.
E911 self-service added to the REST API
Registering an emergency address and enabling E911 on a US phone number is now self-service, and no longer requires a support ticket.
Create a carrier-validated address, assign it to the number, then poll until e911_status reads active.
To get started, read the E911 guide and review the API reference.
Six new text-to-speech engines
You can now use the Groq, Mistral, Grok, Smallest.ai, Fish Audio, and Speechify TTS engines on the SignalWire platform.
playai-turbo is now groq-tts
Groq retired PlayAI in December 2025.
The engine is renamed groq-tts and repointed at Canopy Labs’ Orpheus, defaulting to canopylabs/orpheus-v1-english.
Arabic voices are available by selecting the canopylabs/orpheus-arabic-saudi model, or its shorthand arabic.
See Groq for the voice list.
Rime's arcana model is deprecated, falls back to coda
Rime retired its arcana model.
Applications that invoke it will fall back to coda, which shares voice names.
Additionally, the coda model now uses the same sampling parameters as other Rime models:
repetition_penalty, temperature, top_p, and max_tokens.
See Rime.