Deepgram

Deepgram has two new models for voice agents. Think-Live handles when an agent listens, waits, or speaks, and is generally available in the Voice Agent API. Think-Fast handles tools and replies, and is in early access on the LLM API.

Daily, the company behind Pipecat, is also joining Deepgram. Pipecat stays open source. Deepgram previewed a voice-agent platform built with it, with release timing still to come.

Deepgram no longer supports the dg-token header for verifying callbacks. If your server checks it, switch to Basic Auth by putting credentials in the callback URL and rejecting requests without the expected credentials.

Basic Auth is required for Text-to-Speech and Text Intelligence callbacks, and recommended for Speech-to-Text.

Speech-to-Text also lets you send a random secret through Extra Metadata.…

Deepgram's Flux TTS adds 41 languages and 230 voices on the same streaming and batch endpoints. English voices are unchanged.

German, Dutch, French, Hindi, Italian, Japanese, Portuguese, Russian and Spanish include text normalization. The other 32 languages are in Early Access without it: numbers, dates and abbreviations are read as written, so write them out in words.

Choose a voice from the voice catalog and pass its flux-{voice}-{language} model string as model when calling /v2/speak.…

Deepgram's numerals parameter now formats Japanese numbers on Nova-3 multilingual, Nova-3 Medical multilingual and Flux Multilingual. Japanese formatting already worked on monolingual requests.

Add numerals=true to enable it; it's off by default. On multilingual models, numbers are formatted only in supported languages. Hindi remains unsupported.

Original source

Deepgram's Nova-3 multilingual model now transcribes 55 languages. Existing model=nova-3&language=multi requests automatically use it on the same /v1/listen endpoints for pre-recorded and streaming audio.

The optional languagehint parameter biases detection toward expected languages without excluding others. Repeat it for each language, such as languagehint=en&languagehint=es. It applies only with language=multi, isn't supported by Nova-2, and an unsupported language code fails the request…

Deepgram’s self-hosted release 261020 changes the Flux speech-to-text runtime: it now loads only .dgv2 Flux model files. If you enable Flux with only an older .dg file present, Engine exits at startup with “Can't find Flux model”.

The Engine image also requires NVIDIA driver version 580 or later, using the open kernel module flavor.

All four self-hosted services now have FIPS 140-3 image variants with the -fips tag suffix.…

Deepgram now confirms whether a mid-stream Configure update took effect on a /v1/listen stream. Each Configure message gets exactly one ConfigureSuccess or ConfigureFailure response.

When you replace the stream's keyterms, ConfigureSuccess lists the terms now in effect. A failed update applies none of the message, and the stream stays open with its previous settings.

Keyterm updates require Nova-3 and must stay within the 500-token limit.…

Deepgram's Nova-3 keyterms can now be updated mid-stream on every regional endpoint. Configure messages on /v1/listen work in the EU, Australia and India, as well as on the global endpoint. Deepgram says they behave the same across all of them.

To switch regions, change the base URL; the API keys and SDKs stay the same. Regional requests stay within that region and fail if it's unavailable, rather than falling back elsewhere.

Original source

Deepgram CLI v0.3.3 lets you play generated speech straight from dg speak. Add --play to send audio to ffplay, afplay, paplay, or aplay, and combine it with -o to save and play in one command. Use --list-voices to see the available voices and their languages.

The release also changes dg skills to install Deepgram skills v1.7.0 from a pinned, hash-checked bundle. Installation and removal affect only skill folders the CLI owns; upgrade cleanup removes only files it can prove an earlier 0.3.x…

Deepgram skills v1.7.0 refreshes all 14 skills with September’s API, SDK, and CLI guidance for coding assistants such as Claude Code, Cursor, and Codex.

The updates cover Flux TTS pause markers and Early Access pronunciation overrides, mid-stream numerals in Flux STT, and reusable Voice Agent configurations. They also correct SDK support guidance and add the EU, Australia, and India regional endpoints.

Install with npx skills add deepgram/skills or dg skills.

Original source