Deepgram

Deepgram shipped Flux TTS, a speech synthesis model built for live agent conversations. Its stated difference from typical TTS is conversation-level context: delivery of a line is conditioned on what came before, rather than each utterance being synthesized on its own.

Flux Multilingual is live: real-time conversational speech to text for voice agents across 10 languages, with turn detection and mid-sentence code switching in a single model. The pitch is deploying once rather than stitching language detection and routing in front of separate models.

Nova-3 speech to text picks up Thai, Cantonese Traditional, Mandarin in simplified and traditional script, and Gujarati, alongside accuracy improvements for Bengali, Marathi, Tamil and Telugu. The additions cover both streaming and batch workflows.

Together AI now hosts Deepgram speech to text natively, putting transcription, a Together-hosted LLM and Aura-2 TTS behind one platform. Deepgram cites sub three second round trips on its reference builds. The appeal is fewer vendors and one integration surface for voice agent prototypes.

Deepgram speech-to-text is now available natively on Together AI, letting teams co-locate STT, LLM, and TTS on one platform behind a single API and bill so audio does not leave the environment to be transcribed.

Deepgram's JavaScript SDK v5.0.0 is stable on npm: regenerated codegen with stronger typing, one package covering browser and Node, stable Voice Agent v1 interfaces, Nova-3 and Flux support, and DEEPGRAM_ACCESS_TOKEN auth. v4 users need the migration guide.

Deepgram's self-hosted deployment now supports a FIPS 140-3 compliant configuration across API, Engine and License Proxy, with FIPS-encrypted models on Chainguard's validated crypto module. It is a configuration flag on the same Nova models, and the post covers TLS 1.3 cutover details.

Deepgram released three developer surfaces in April: a dg CLI for terminal speech workflows, a dg MCP server for coding agents, and a skills package that preloads Deepgram context on install so agents stop guessing at the API.

Deepgram's Voice Agent API now runs on NVIDIA Nemotron, with a joint writeup covering the pipeline architecture, latency benchmarks, and how to deploy the same agent across cloud, VPC, and on-premises environments.

Deepgram's Flux Multilingual runs 10 languages in a single real-time stream with automatic language detection, native code switching, and a language_hint parameter for biasing, replacing the detection service, per-language models, and routing layer most multilingual stacks assemble.

Deepgram argues most voice agent misbehavior starts in the audio path rather than the prompt or the LLM. A clean transcript still leaves agents reacting to background noise or their own echo, so agent speech models need echo cancellation, overlap detection, and primary speaker identification, not ju

Version 5.0.0 of Deepgram's JavaScript SDK is stable on npm: regenerated from codegen for stronger typing, native in both browser and Node, stable Voice Agent v1 interfaces covering Nova-3 and Flux, and auth moved to DEEPGRAM_ACCESS_TOKEN. It breaks v4, with a migration guide alongside.

Deepgram improved Nova-3's Portuguese transcription accuracy across Brazilian and European variants. It is available now under model=nova-3 with the pt, pt-BR, and pt-PT language codes, testable in the playground.

Deepgram is porting Nova-3 to run locally on the Qualcomm Hexagon NPU in Snapdragon X Series PCs. Transcription then happens on the machine with no cloud round trip, cutting latency and keeping audio off the network entirely.

Deepgram shipped a Browser Agent SDK as four composable npm packages: a drop-in widget, React components, React hooks, and a framework-agnostic core. Reconnection, audio buffering, playback-aware mode tracking, KeepAlive, Silero VAD, and typed events behave the same at every layer.

Deepgram's Australian region is live in Sydney, keeping audio, transcripts, and synthesized speech inside the country. The API surface and models are unchanged, so residency-bound teams migrate by pointing at a different base URL rather than standing up a self-hosted deployment.

Deepgram extended Flux to multilingual restaurant ordering across English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, and Dutch. The target is drive-thru and support flows where callers switch out of English mid-interaction.

Deepgram's second-generation batch diarization is live for pre-recorded audio, selected through a new diarize_model parameter. Deepgram reports human evaluators preferred its speaker labeling 3.3x over the prior system, with better attribution on real-world recordings.

Deepgram's Flux streaming ASR now accepts a Configure message mid-connection, letting a voice agent swap keyterms and loosen or tighten turn-detection thresholds as a call moves between reading back a PIN, reviewing drug names, and open conversation.

Deepgram's Voice Agent API now lists Claude Sonnet 5 among its selectable reasoning models, chosen in agent configuration rather than through code changes. The developer changelog carries the full set of supported models alongside it.

Deepgram rolled out monolingual quality upgrades to Nova-3 across 50 languages, improving transcription accuracy and keyterm prompting in both batch and streaming speech to text, and added Afrikaans, Armenian, Georgian, Nepali, Punjabi, Assamese, Pashto and Mongolian.

Deepgram's self-hosted 260630 release adds per-word timestamps to Flux and corrects streaming timestamp accuracy, plus Spanish number formatting fixes and broader number and currency formatting. The self-hosted models endpoint returns a restructured response.

A Deepgram and Twilio session on what changes when a voice agent moves from a web page to a phone line: where latency accumulates between carrier and model, how codecs degrade recognition, handling barge-in and callers who trail off, answering machine detection, and outbound carrier trust rules.