Deepgram adds two voice models and welcomes Daily
Think-Live runs the conversation and is generally available. Think-Fast does the work and is in early access. Daily, the company behind Pipecat, is joining Deepgram.
Every Deepgram update in one feed — release notes, features and changes, straight from the source.
52 updates
Think-Live runs the conversation and is generally available. Think-Fast does the work and is in early access. Daily, the company behind Pipecat, is joining Deepgram.
Deepgram callbacks no longer support dg-token verification; use Basic Auth for STT, TTS and Text Intelligence, or an Extra Metadata secret for STT.
Requests using model=nova-3&language=multi get the new model automatically, and an optional language_hint parameter steers detection toward expected languages.
Numerals now formats Japanese on Nova-3 and Nova-3 Medical multilingual and Flux Multilingual, while Hindi stays unsupported.
Flux TTS adds voices in languages from Bengali to Vietnamese on the same endpoints, with text normalization in nine and Early Access for the other 32.
Deepgram's 261020 self-hosted images add FIPS 140-3 variants and a new Flux runtime that no longer loads old .dg model files.
Every Configure message on a Nova-3 /v1/listen stream now gets a ConfigureSuccess or ConfigureFailure reply, with the keyterms in effect.
Nova-3 Configure messages on /v1/listen now work on the EU, Australia and India endpoints, and behave the same as on the global endpoint.
Python developers can update live Nova-3 streams without reconnecting and exchange experimental custom Think-provider messages.
Developers can update Nova-3 keyterms mid-stream and exchange experimental custom Think-provider messages in Voice Agent sessions.
Go developers get typed Voice Agent function calls and management endpoints for reusable configurations and template variables.
Listen V1 WebSocket connections now accept an omitted model, with or without keyterms.
The release adds text analysis and caption helpers, with breaking changes to usage fields and custom base URL handling.
Streaming fixes prevent hangs and expose terminal transport errors, while sends on ended /v1/listen sessions now return Err.
New flags play generated speech and list voices, while skill installs use a pinned, hash-checked v1.7.0 bundle.
Java developers can update Nova-3 keyterms mid-stream, handle typed server errors and exchange custom Think-provider payloads.
All 14 skills refresh API, SDK, and CLI guidance for coding assistants including Claude Code, Cursor, and Codex.
The SDK updates add mid-stream Numerals controls, Flux TTS pause markers and Early Access pronunciation overrides.
Exceeding 500 tokens stops Nova-3 transcription without an Error, then closes the stream after about 30 seconds.
Hosted users receive the improvements automatically; self-hosted customers must request updated models from their account representative.
Hosted and self-hosted customers can use the API for both pre-recorded medical audio and real-time streaming.
The updates apply automatically to existing API requests; self-hosted customers must request updated models from their account representative.
Nova-3 streaming users can replace keyterms with a Configure message on the global endpoint; EU, Australia and India endpoints aren't supported yet.
Flux TTS audio is now watermarked, and self-hosted Engine will not start with Flux TTS enabled until the watermarker file and UUID are configured.
Flux TTS adds inline pauses on batch requests and Early Access IPA pronunciation overrides on batch and streaming, with errors for invalid combinations.
Voice app developers can toggle digits versus words on a live Flux STT stream, on the English model and eight languages of the multilingual one.
Nova-3 transcription improves for German (Switzerland), Portuguese, Lithuanian and Flemish, with no API changes required.
Nine Nova-3 language models get quality updates for batch or streaming transcription, live now with no API changes required.
Deepgram says client-side Silero VAD was documented but never shipped in the Browser Agent SDK; barge-in runs server-side instead.
Deepgram fixes a docs example that silently ignored the token lifetime developers tried to set for Voice Agent API calls.