Ce que Deepgram a publié en juin 2026

Deepgram Self-Hosted June 2026 release adds Flux word-level timestamps and model metadata changes

Deepgram’s June 2026 self-hosted release (260630) introduces Flux word-level timestamps, fixes Spanish number formatting, expands number/currency formatting, and improves streaming timestamp accuracy.…

Changelog | Product Enhancements

Deepgram integrates STT/TTS with Amazon Connect and Lex

Deepgram now supports integration with Amazon Connect and Amazon Lex, allowing organizations to replace AWS Transcribe and Polly with Deepgram's STT and TTS models without altering existing Connect fl…

Introducing Deepgram’s Voice Agent API

Deepgram introduced the Voice Agent API, a unified voice-to-voice API enabling enterprises to build LLM-powered AI agents for customer support and order taking. The API leverages industry-leading spee…

Radiology speech recognition: dictation accuracy for diagnostic workflows

Deepgram introduced Nova-3 Medical, a radiology-specific speech-to-text API addressing up to 22% error rates in general-purpose models. Nova-3 Medical achieves a 6.79% keyword error rate by handling d…

What Is Automatic Speech Recognition and How Does It Work?

Deepgram’s product marketing team published a comprehensive guide on automatic speech recognition (ASR) in 2026, detailing four model architectures (streaming conformers, encoder-decoder transformers,…

Top APIs for Building Programmable Voice Agents: What the Stack Looks Like in Production

Your misbehaving agent is likely an audio quality problem

Deepgram argues that misbehaving voice agents typically result from poor audio quality rather than model or prompt flaws. Background noise causes transcription errors and false interruptions, while ha…

Restaurants stack every audio quality problem at once, and Deepgram is fixing it

Deepgram addresses restaurant-specific audio challenges by introducing primary speaker identification in its speech-to-text model, using rolling enrollment windows to distinguish the customer’s voice …

Conversational intelligence for enterprise sales: CRM pipelines, coaching, and revenue attribution

Deepgram’s blog post argues that conversation intelligence (CI) for sales teams fails unless the underlying speech-to-text (STT) layer delivers high accuracy on entity extraction (names, companies, co…

Speaker Labels in STT Output: Formats, Configuration, and Accuracy for Multi-Speaker Audio

Deepgram published a technical guide explaining how speaker labels work in their Speech-to-Text API, comparing output formats across providers, and detailing configuration parameters that affect diari…

Deepgram adds runtime speed and pronunciation controls to Aura-2 TTS

Deepgram’s Aura-2 TTS voices now support runtime speed (0.7x–1.5x) and pronunciation overrides via IPA notation in English and Spanish. Controls apply to batch, streaming, and Voice Agent APIs, with s…

Webinar en Vivo: Desarrollar Voice Agents listos para producción con Deepgram

Deepgram announced a technical webinar on July 7, 2026, focused on building voice agents ready for enterprise deployment in Spain. The session will cover real-time voice recognition, synthesis, full-s…

Deepgram's Australia Endpoint Is Now Generally Available

Deepgram’s Australia endpoint is now generally available, enabling fully managed voice AI workloads with data processed and stored entirely within Australia. The endpoint supports speech-to-text, text…

Deepgram launches Australia endpoint for Speech-to-Text, TTS, Voice Agent, and Text Intelligence APIs

Deepgram announced the general availability of its Australia endpoint (api.au.deepgram.com) supporting Speech-to-Text, Text-to-Speech, Voice Agent, and Text Intelligence APIs. Customers can now proces…

Real-time speech analytics: Building live dashboards on your voice API stack

Deepgram published a technical guide on constructing real-time speech analytics dashboards using their voice API, covering pipeline stages (capture to UI push), latency control, and compliance (HIPAA,…

Production-Ready Voice Agents: What Separates a Demo from an Enterprise Deployment

Deepgram explains why enterprise voice agent deployments fail despite demo success, highlighting latency under load, accuracy on noisy audio, bundled pricing predictability, and compliance needs like …

Conversation intelligence for voice agents: post-call and real-time analytics

Deepgram updated its Audio Intelligence API to support both real-time and post-call analytics for voice agents, enabling structured signals like intent, sentiment, and topic detection. Real-time featu…

Medical transcription in 2026: how AI replaces legacy services at clinical scale

Deepgram’s product marketing article outlines 2026’s medical transcription landscape, comparing AI speech engines, finished AI scribes, and human BPO services. It highlights cost, accuracy, compliance…

Deepgram introduces UpdateListen for on-the-fly voice agent tuning

Deepgram launched UpdateListen, enabling real-time adjustment of voice agent settings like end-of-turn thresholds, keyterms, and language hints without restarting sessions. The feature supports partia…

Deepgram Self-Hosted June 2026 release adds Persian profanity filtering and streaming diarization

Deepgram’s June 2026 Self-Hosted release (260611) introduces Persian profanity filtering, English redaction for Flux streaming, a streaming diarization model parameter, and Text-to-Speech transcoding.…

Deepgram adds reusable agent configurations and template variables

Deepgram introduced reusable agent configurations via its API, allowing users to store and reference agent setups by UUID instead of resending full configurations per session. This supports per-custom…

Deepgram adds NVIDIA as LLM provider for Voice Agent API

Deepgram’s Voice Agent API now supports NVIDIA as an LLM provider, offering the `nemotron-3-nano-30B-A3B` model in the Standard pricing tier. The integration simplifies agent configuration by allowing…

Deepgram adds thought_signature for Gemini and volume control for Cartesia TTS

Deepgram’s Voice Agent API now includes an optional `thought_signature` field to improve function calling performance for Gemini 3.0/3.1 models, addressing prior degradation. It also adds a `volume` p…

Deepgram Nova-3 adds Chinese language support

Deepgram’s Nova-3 model now supports Simplified and Traditional Chinese (language codes `zh`, `zh-CN`, `zh-Hans`, `zh-TW`, `zh-Hant`). Users can access these via the `model="nova-3"` parameter in API …

Introducing Batch Diarization V2

Deepgram introduced Batch Diarization V2, a major upgrade to its speaker labeling system for pre-recorded audio, improving attribution accuracy and reducing errors by up to 3.3X in human evaluations. …

Live Webinar: Produktionsreife Voice Agents mit Deepgram entwickeln

Deepgram’s DACH team will host a technical webinar on June 25, 2026, demonstrating how to develop production-ready voice agents for German enterprise deployments. The session will cover real-time spee…

Accent Variability in Restaurant Voice AI: Why the ASR Layer Makes or Breaks Multi-Region Ordering

Deepgram highlights how accent variability in ASR layers undermines multi-region restaurant voice ordering accuracy, citing McDonald's pilot failures. It introduces techniques like Keyterm Prompting (…

Enterprise restaurant brands deserve frontier voice AI models

Deepgram introduced Deepgram for Restaurants, a voice AI platform designed for enterprise-scale restaurant brands with complex menus and POS systems. The solution provides custom fine-tuning, noise fi…

What Is a HIPAA Business Associate Amendment and When Does Voice AI Need One?

Proposed HIPAA Security Rule updates (effective 2026) and voice AI adoption are creating gaps in existing Business Associate Agreements (BAAs). Real-time audio processing, transcription, and subcontra…

Named Entity Recognition for Voice: Extracting Structure from Transcripts

Deepgram’s article highlights that NER models lose 20–27 F1 points on raw ASR output due to formatting, casing, and error cascades. It compares pipeline, LLM, and joint-model architectures, emphasizin…

How to Integrate Voice Agents with Salesforce for Enterprise Service and Sales

Deepgram’s guide outlines two approaches to integrate voice agents with Salesforce: native Agentforce Voice or external STT APIs via Salesforce’s Telephony Integration API. The article emphasizes the …

Running an AI Call Center Voice Agent in Production: An Orchestration Playbook

Deepgram published a production playbook for AI call center voice agents, detailing latency budgets, failure modes, cost modeling, and monitoring metrics. It highlights the orchestration layer as the …

Voice Agent API Architecture: What Bundled Pricing Changes for Builders

Deepgram’s article details how bundled voice agent APIs (packaging STT, LLM orchestration, and TTS) alter architecture by reducing middleware, simplifying scaling, and imposing customization limits. I…

Deepgram Delivers Private Voice AI to Regulated Industries with On-Premises Deployments Powered by Fortanix Confidential AI and NVIDIA Confidential Computing

Deepgram partnered with Fortanix and NVIDIA to enable enterprises to run voice AI in private environments using Fortanix Confidential AI and NVIDIA Confidential Computing. This integration protects se…