Ce que Deepgram a publié en juillet 2026

Voice Agents vs. Automation Platforms: Where Workflow Tools End and Conversational AI Begins

Deepgram publishes a blog post differentiating voice agents from automation platforms, positioning conversational AI as a distinct category beyond traditional workflow tools.

Deepgram Voice Agent gains mid-session model switching, latency reporting, and Google Enterprise Agent API support

Deepgram’s Voice Agent now supports switching speech-to-text models and languages mid-session via `UpdateListen`, adds `stt_latency` to `LatencyReport` for real-time performance tracking, and enforces…

Deepgram Self-Hosted Is Now FIPS 140-3 Compliant

Deepgram now offers a FIPS 140-3-compliant configuration for its self-hosted speech models, enabling deployment in federal, defense, and public safety environments without accuracy tradeoffs. The comp…

The Real Cost of Voice AI Deployment: Dedicated vs. Self-Hosted vs. Cloud

Deepgram’s blog post breaks down the true costs of cloud, dedicated, and self-hosted voice AI deployment, highlighting hidden labor and infrastructure expenses. It argues that managed cloud often wins…

Pricing change detected for Deepgram

Pricing updated for Deepgram: - Pay-As-You-Go: promotion changed — Free $200 Credit - New tier: Custom (Custom pricing) - Removed: Developers & Startups tier

Deepgram Self-Hosted July 2026 release adds FIPS 140-3, .dgv2 model format, and formatting improvements

Deepgram’s July 28, 2026 self-hosted release (260728) introduces FIPS 140-3 compliant container images in GA, a new encrypted .dgv2 model file format, and enhanced diarization metadata. The update als…

Pricing updated for Deepgram: - Pay-As-You-Go: promotion changed — Free $200 Credit then pay-as-you-go - New tier: Developers & Startups ($4000/yr) - Removed: Custom tier

Per-minute vs. per-character vs. concurrency: how voice AI pricing actually works

Deepgram published a guide breaking down voice AI pricing models—per-minute, per-character, concurrency, and volume—highlighting how each measures different cost drivers and where hidden costs typical…

The platform tax: what you're actually paying for a managed voice agent

Deepgram’s product marketing team published a detailed analysis of hidden costs in managed voice agent platforms, highlighting platform tax, integration tax, and observability tax. The article compare…

Deepgram adds numerals support and Aura-2 TTS voices to JavaScript, Python, and Java SDKs

Deepgram introduced `numerals` parameter support across JavaScript, Python, and Java SDKs to convert spoken numbers into numerical format during Flux WebSocket connections. All three SDKs (v5.7.0, v7.…

Deepgram Delivers Real-Time Voice AI at the Edge for Use withSnapdragon

Deepgram announced the optimization of its Nova-3 speech-to-text model to run on Qualcomm’s Snapdragon X Series processors, enabling real-time, on-device voice AI without cloud dependency. The integra…

Deepgram releases Nova-3 models for Kannada, Telugu, and Greek

Deepgram launched Nova-3 monolingual models for Kannada, Telugu, and Greek, expanding its language coverage for both batch and streaming speech recognition workloads. The updates are live for all user…

Deepgram Flux now supports numeral formatting

Deepgram added a new `numerals` feature to its Flux speech-to-text models, converting written numbers (e.g., 'nine hundred') to numerical format (e.g., '900'). The feature is available for Flux Englis…

Deepgram launches Flux TTS with streaming-first voice agent pipeline

Deepgram introduced Flux TTS, a new speech synthesis model and API designed for voice agent pipelines rather than one-shot text-to-audio conversion. It supports both real-time (WebSocket) and batch (R…

Deepgram releases improved Nova-3 models for 10+ languages

Deepgram launched improved Nova-3 monolingual models for 10 languages, including Ukrainian, Catalan, Turkish, and French, enhancing transcription accuracy for both batch and streaming workloads. Users…

Deepgram Self-Hosted July 2026 release adds number formatting and no-audio timeout close reason

Deepgram released Self-Hosted version 260714 with decimal/percentage formatting improvements (trailing zeros preserved) and Malay spoken-number conversion. Streaming sessions now return a dedicated `n…

Conformer Models: ASR Accuracy Benchmarks and Trade-offs

Deepgram’s product marketing team published a technical deep dive on Conformer models, highlighting their hybrid convolution-self-attention architecture, benchmark performance (e.g., 2.1% WER on Libri…

Connecting TTS to Twilio: The Complete Guide to Audio Encoding, Streaming, and Production Gotchas

Deepgram published a technical guide explaining how to configure their TTS API for compatibility with Twilio Media Streams, including required audio encoding (8kHz mu-law), WebSocket framing, and barg…

Is ElevenLabs Real-Time? What Developers Need to Know

Deepgram's analysis argues ElevenLabs' 75ms model inference latency excludes network and application overhead, with production deployments facing 180ms gaps in measured TTFB. Concurrency limits (max 3…

Deepgram Voice Agent API vs OpenAI Realtime API: Building Voice Agents in 2026

A detailed comparison of Deepgram’s Voice Agent API and OpenAI’s Realtime API highlights critical differences in pricing models, architecture, latency, concurrency limits, and deployment control. Deep…

Deepgram Vs. Rev AI Vs. Whisper: Latency, Accuracy, and Scalability Tested

Deepgram published a comparative analysis testing their speech recognition technology against Rev AI and Whisper, focusing on latency, accuracy, and scalability metrics.

Deepgram adds interrupt mode to InjectAgentMessage and full LatencyReport support

Deepgram expanded its voice agent API with two key updates: a new `interrupt` behavior for `InjectAgentMessage` that allows agents to immediately speak or interrupt ongoing speech, and full support fo…

When the docs write the code: migrating my Speech-to-Text playground to Python SDK v6

Deepgram’s v6 Python SDK is fully generated from the API schema, replacing hand-built WebSocket clients and requiring extensive code changes. A Senior Applied Engineer rewrote a customer-facing tool f…

Voice recognition in healthcare: compliance, accuracy, and deployment realities

Deepgram argues benchmark Word Error Rate (WER) overstates clinical voice AI accuracy, highlighting medical-entity error rates and real-world noise as key differentiators. The post details HIPAA compl…

AI Voice Agent for Benefits Verification: Automating the Eligibility Call

Deepgram introduced an AI voice agent that automates eligibility and benefits verification calls, reducing manual processes that cost $14.32 per transaction. The agent handles both electronic 270/271 …

AI Phone Agent for Patient Intake: What Happens Before the Appointment

Deepgram introduced an AI phone agent designed to automate patient intake workflows, including demographics collection, insurance ID verification, and visit reason documentation. The system relies on …

Automated Patient Reminders: How AI Voice Agents Reduce No-Shows

Deepgram introduced AI voice agents that automate patient reminders, enabling real-time rescheduling, waitlist backfill, and multilingual outreach within the same call. The system integrates with EHRs…

Deepgram Flux adds word-level timestamps in API responses

Deepgram’s Flux model now returns word-level timestamps in its speech-to-text responses, including start and end times for each word in the transcript. The change is additive and backward-compatible, …

Deepgram adds Claude Sonnet 5 to Voice Agent API, deprecates Sonnet 4

Deepgram introduced `claude-sonnet-5` as a managed Anthropic LLM in its Voice Agent API, offering improved reasoning and conversational quality for voice agents. Concurrently, the older `claude-sonnet…