What Deepgram shipped in April 2026

Deepgram introduces Aura-2 voice controls for speed and pronunciation adjustments

Deepgram added Aura-2 voice controls to its text-to-speech API, enabling fine-grained adjustments to speaking speed (0.7x–1.5x) and pronunciation overrides using IPA notation. Both REST and WebSocket …

Deepgram introduces InjectAgentMessage for mid-conversation agent injections

Deepgram added the `InjectAgentMessage` feature to its Voice Agent, allowing servers to inject agent statements during live conversations. The feature supports two behaviors: `default` (waits for sile…

Deepgram Voice Agent adds support for multiple LLM providers and models

Deepgram’s Voice Agent now supports a broader range of LLM providers and models, including OpenAI’s latest GPT-5 series, Anthropic’s Claude 4 models, Google’s Gemini 3 series, Groq’s GPT OSS 20B, and …

Deepgram expands Voice Agent LLM model support with NVIDIA Nemotron and additional providers

Deepgram’s Voice Agent now supports NVIDIA’s Nemotron-3-Nano-30B-A3B model under the `nvidia` provider type, alongside existing OpenAI, Anthropic, Google, Groq, and AWS Bedrock options. The update als…

Deepgram expands Voice Agent TTS options with Cartesia and third-party providers

Deepgram’s Voice Agent API now supports multiple TTS providers beyond its native models, including managed Cartesia TTS and third-party options like OpenAI, ElevenLabs, Amazon Polly, and Cartesia. Use…

Deepgram introduces Flux and Nova-3 speech models with turn detection and multilingual support

Deepgram launched Flux, a conversational ASR model with model-native turn detection for voice agents, and Nova-3, a high-accuracy general-purpose model with 54.2% lower WER in streaming and 47.4% in b…

Deepgram launches MCP server for AI coding tools

Deepgram introduced a built-in MCP server in its `dg` CLI, enabling AI coding assistants like Claude Code, Cursor, and Windsurf to directly access Deepgram APIs for transcription, speech synthesis, te…

Deepgram introduces multi-agent architecture for voice agents with specialized agent phases

Deepgram released a multi-agent architecture for voice agents, replacing single-agent systems with a phased approach using specialized agents (Qualifier, Advisor, Closer) for focused tasks. The system…

Deepgram introduces reusable agent configurations with template variables for Voice Agent API

Deepgram launched reusable agent configurations via API, allowing users to store and reference agent blocks by UUID instead of repeating full configurations in Settings messages. The feature supports …

Deepgram CLI installation methods expanded with Homebrew and pipx support

Deepgram updated its CLI installation process with expanded support for Homebrew, pipx, and uv, alongside existing script-based and pip methods. Homebrew now auto-installs dependencies like ffmpeg, si…

Deepgram Self-Hosted April 2026 release adds Gujarati, Aura-2 controls, and Voice Agent improvements

Deepgram’s April 2026 Self-Hosted release (260430) introduces Nova-3 Gujarati support, Aura-2 speed and pronunciation controls, multilingual numeral formatting for Nova-3, and Voice Agent enhancements…

Deepgram Self-Hosted April 2026 release adds Gujarati, Aura-2 controls, and numeral formatting

Deepgram’s April 2026 self-hosted release (260430) introduces Nova-3 support for Gujarati, Aura-2 speed and pronunciation controls requiring an updated voice-pack, multilingual numeral formatting for …

Pricing change detected for Deepgram

Pricing updated for Deepgram: - Custom: allowance changed from $4K+ / year to For businesses with large volumes, data or deployment requirements, or support needs - New tier: Free ($0/mo — promo: $200…

Cartesia deprecates speed and emotion controls in TTS API

Cartesia deprecated its speed and emotion controls feature for text-to-speech, previously available via API and playground. The feature was experimental and subject to breaking changes, with controls …

Flux Multilingual Technical Deep Dive: Multilingual Speech-to-Text Without the Routing Mess

Deepgram introduced Flux Multilingual (flux-general-multi), a single real-time streaming model handling 10 languages with automatic detection and code-switching. A new `language_hint` parameter biases…

How to Master Real-Time Transcription for Any Workflow

Deepgram released a six-phase roadmap for implementing real-time transcription from proof-of-concept to production, covering requirements, benchmarking, integration, accuracy tuning, scaling, and comp…

Deepgram vs Speechmatics vs Rev AI: Scale Comparison

A detailed comparison of Deepgram, Speechmatics, and Rev AI highlights architectural differences in concurrency limits, pricing models, latency, and compliance. Deepgram leads in managed real-time sca…

Introducing Flux Multilingual: One Conversational Speech Model for Global Voice Agents

Deepgram introduced Flux Multilingual, a single conversational speech model supporting 10 languages (English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, Dutch) with monoli…

Deepgram Launches Flux Multilingual: The World’s First Multilingual Conversational Speech Recognition Model

Deepgram released Flux Multilingual, a general-availability conversational speech recognition model supporting 10 languages with monolingual-grade accuracy and real-time language switching. The model …

Deepgram adds GPT-5.5 LLM and Cartesia TTS speed control

Deepgram introduced OpenAI’s GPT-5.5 as a managed LLM in its Voice Agent API and added speed control for Cartesia TTS, supporting preset values and numerical tuning. The Llama Nemotron Super 49B model…

Deepgram introduced OpenAI's GPT-5.5 as a managed LLM in its Voice Agent API and added speed control for Cartesia TTS. The Llama Nemotron Super 49B model was removed due to poor performance.

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

Deepgram launches Voice Agent API with real-time LLM integration

Deepgram introduced a new Voice Agent API enabling real-time conversational AI with integrated LLM capabilities. The API supports multi-language audio input/output, configurable models (e.g., OpenAI's…

Deepgram CLI text-to-speech tool expands voice options and streaming

Deepgram’s CLI text-to-speech tool now supports additional output formats (WAV, MP3, FLAC), voice selection via `--voice` and `--list-voices`, language selection, and low-latency streaming via WebSock…

Deepgram CLI adds text intelligence features for sentiment, topics, summarization, and intent detection

Deepgram expanded its CLI with text intelligence capabilities including sentiment analysis, topic detection, summarization, and intent recognition. Users can now analyze documents, URLs, or piped inpu…

Deepgram CLI adds shell completion for bash, zsh, and fish

Deepgram introduced shell completion scripts for its CLI, enabling tab-completion for commands like 'dg listen' and subcommands such as '--mic' or '-o json'. Users can generate and install completions…

Deepgram CLI introduces plugin system for extensibility

Deepgram’s CLI now supports a plugin system allowing users to install, update, and uninstall Python-based plugins that add new commands. Plugins run in isolated environments, can access Deepgram confi…

Eliezer testing

Deepgram Nova-3 adds Gujarati language support

Deepgram’s Nova-3 model now supports Gujarati (language codes `gu`, `gu-IN`), expanding its multilingual capabilities. Users can access the model by specifying `model="nova-3"` and the Gujarati langua…

Deepgram’s Nova-3 speech-to-text model now supports Gujarati (language codes `gu`, `gu-IN`). Users can access this by setting `model="nova-3"` and the relevant language code in API requests.

Deepgram vs Google vs Azure: Which Cloud Provider Wins at STT?

Deepgram published a comparative analysis of its speech-to-text (STT) service against Google Cloud and Azure, focusing on real total cost of ownership (TCO), latency, compliance, and deployment flexib…

Deepgram

Deepgram published a practitioner-level guide on building and deploying Voice AI agents, focusing on real-time distributed systems that coordinate speech, reasoning, and audio under strict latency con…

Deepgram vs Speechmatics vs AssemblyAI: Finding the Right Fit for Your Team

Deepgram, Speechmatics, and AssemblyAI were compared across latency, pricing, language support, and deployment for speech-to-text (STT) workloads. Deepgram excels in real-time voice agent infrastructu…

Large Vocabulary Speech Recognition Demystified

Deepgram introduced Keyterm Prompting, a runtime vocabulary customization feature in its Nova-3 model, to address out-of-vocabulary (OOV) terms in large vocabulary speech recognition. Keyterm Promptin…

Deepgram Self-Hosted 260416 adds Flux Multilingual STT with code-switching

Deepgram released Self-Hosted version 260416, introducing Flux Multilingual for real-time multilingual conversational speech-to-text (STT) with code-switching support across 10 languages. The update r…

Deepgram Self-Hosted April 2026 release adds Flux Multilingual STT

Deepgram’s April 2026 Self-Hosted release (260416) introduces Flux Multilingual, enabling real-time multilingual conversational speech-to-text with code-switching for 10 languages. The feature require…

Flux Multilingual & Language Prompting

Deepgram introduced Flux Multilingual (`flux-general-multi`), a single model supporting 10 languages with near-monolingual accuracy when language hints are provided. The model auto-detects languages i…

Configure: Real-Time Dynamic Adjustments for Flux Streaming ASR

Deepgram introduced a real-time `Configure` control message for its Flux streaming speech recognition system, enabling mid-stream adjustments to key recognition parameters without disconnecting. This …

Getting Started with Flux

Deepgram launched Flux, the first conversational speech recognition model designed specifically for voice agents, moving beyond traditional speech-to-text (STT) by understanding conversational flow an…

Deepgram Launches Real-Time Conversational Speech Recognition API with Turn Detection

Deepgram introduced a new real-time conversational speech recognition API endpoint, `/v2/listen`, designed for natural voice conversations with contextual turn detection. The API enables developers to…

The teams we empower

Deepgram highlights its Voice AI platform's adoption across diverse industries, including startups, NASA, and contact centers like Five9. The company emphasizes its ability to process millions of audi…

Articles

Deepgram has announced a new funding round and the acquisition of Of.One, marking a significant expansion in its voice AI capabilities. The company highlights its growing influence in the voice AI eco…

The Voice AI Economy isPowered by Deepgram

Deepgram introduces a unified Voice Agent API that consolidates speech-to-text, text-to-speech, and LLM orchestration into a single interface, reducing complexity, latency, and cost for businesses. Th…

Deepgram Launches Voice Agent API for Real-Time Conversational Voice Agents

Deepgram has launched a new Voice Agent API endpoint, enabling developers to build real-time conversational voice agents using a WebSocket-based interface. The API introduces a bidirectional message s…

Creates a model response for the given chat conversation.

NVIDIA updated the API reference documentation for its Llama 3.3 Nemotron Super 49B v1.5 model, introducing new parameters for text generation control. The changes include support for token generation…

NVIDIA updated the API reference documentation for its Nemotron-3-Nano-30B-A3B-Infer model, clarifying key parameters for text generation. The changes specify the required structure of conversation me

Revolutionizing contact center AI with Deepgram and Five9

Deepgram and Five9 have integrated Deepgram’s Nova-2 automatic speech recognition (ASR) model into Five9’s Intelligent Virtual Agent (IVA) Studio 7 to enhance contact center AI capabilities. The integ…

NASA uses Deepgram to power the next generation of space tech

NASA has adopted Deepgram’s AI Speech Platform, Tailored Speech Models, and Audio Search to address critical challenges in space mission communications. The primary change is NASA’s shift from manual …

Cartesia’s Sonic-3 Adds Volume, Speed, and Emotion Controls

Cartesia has expanded its Sonic-3 text-to-speech (TTS) model with new controls for volume, speed, and emotion, enabling more expressive and customizable speech generation. Users can now adjust these p…

Elevating call center efficiency with Deepgram's Voice AI platform

Deepgram’s advanced speech-to-text (STT) technology has been integrated into MaxContact’s cloud contact center platform to improve transcription accuracy, particularly for mono recorded calls. This in…

Speech Recognition: How It Works and Key Applications

Speech recognition technology converts spoken language into text, with production-grade systems requiring more than benchmark accuracy to handle real-world conditions like noise, accents, and domain-s…

Menu Data is Like a Box of Chocolates: What Happens When You Give Engineers the Keys and a Coke Zero, Part 2

Deepgram introduced an agentic menu integration pipeline to normalize unstructured and chaotic menu data from restaurant POS systems, enabling AI-driven voice ordering. The tool ingests raw POS data, …

Pricing discovered for Deepgram

Pricing updated for Deepgram: - New tier: Pay-As-You-Go (Custom pricing) - New tier: Enterprise (Custom pricing)

Speech Recognition: Models, Challenges, Solutions

Speech recognition has evolved from a multi-stage traditional ASR pipeline to a single neural network model that maps audio directly to text, simplifying the stack but making architecture choice centr…

Today We Start the Next Chapter of the Voice AI Economy

Deepgram has raised a Series C funding round and acquired Of.One, marking a significant milestone in its decade-long journey to dominate the voice AI ecosystem. The company now powers the majority of …

Virtual Event: AI-Powered Outbound Dialing in Healthcare

Deepgram and AWS announced a joint webinar demonstrating an AI-powered outbound dialing architecture for healthcare, specifically targeting patient outreach challenges like clinical trial recruitment,…

Deepgram launches CLI with MCP server for AI coding tools

Deepgram introduced a new CLI tool (`dg`) that unifies transcription, speech synthesis, text analysis, and account management in a terminal command. It also includes an MCP server for integration with…

Flux Multilingual: Conversational STT Now in 10 Languages

Deepgram has launched Flux Multilingual, a single speech-to-text model supporting 10 languages (English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, and Dutch) with convers…

Changelog

Deepgram introduced reusable agent configurations and template variables via API, allowing users to store and reference agent setups by UUID instead of resending full configurations per session. The c…

Reusable Agent Configurations Now Available via Deepgram API

Deepgram introduced reusable agent configurations via its API, allowing users to store and manage agent setups and template variables by UUID instead of resending full configurations per WebSocket ses…

NVIDIA LLM Provider Now Available for Deepgram’s Voice Agent API

Deepgram has added NVIDIA as a supported LLM provider for its Voice Agent API, introducing two new models—`llama-nemotron-super-49B` and `nemotron-3-nano-30B-A3B`—available in the Standard pricing tie…

Deepgram Self-Hosted April 2026 release adds certificate endpoint fix and model canonical_name field

Deepgram released Self-Hosted version 260402, fixing the Engine certificate endpoint path to `/v1/certificates` and adding a `canonical_name` field to the `/v1/models` response in API 1.181.3 and Engi…

Deepgram Self-Hosted April 2026 Release (260402)

Deepgram’s April 2, 2026 Self-Hosted release (260402) introduces two key changes: a fix to the Engine certificate endpoint path to align with other container images and the addition of a canonical_nam…

Deepgram Voice Agent API adds `thought_signature` for Gemini and `volume` parameter for Cartesia TTS

Deepgram’s Voice Agent API has added an optional `thought_signature` field to function call messages, specifically for Google’s Gemini 3.0 and 3.1 models, to address degraded function calling performa…