What Deepgram shipped in August 2026

Deepgram expands Flux TTS speed range from 0.85–1.15 to 0.5–1.5

Deepgram widened the speed parameter range for Flux TTS from 0.85–1.15 to 0.5–1.5 in 0.05 increments, while keeping the default at 1.0. The change applies to streaming WebSocket, batch REST, and Voice…

Deepgram Nova-3 model updates expand multilingual ASR support

Deepgram released improved Nova-3 monolingual models for ten languages (Bulgarian, Croatian, Estonian, Georgian, Italian, Latvian, Lithuanian, Malay, Marathi, Telugu), enhancing transcription quality …

Batch Text-to-Speech Deep Dive: From One Speech Call to a Produced Show

Deepgram published a technical deep dive on a 13-stage batch text-to-speech pipeline enabling the production of long-form audio like podcasts from plain text without SSML or tuning. The pipeline lever…

Deepgram Nova-3 adds Assamese, Mongolian, Pashto and improves Czech, Danish, Swedish, Tagalog, Turkish models

Deepgram expanded Nova-3 with three new languages (Assamese, Mongolian, Pashto) and improved transcription quality for Czech, Danish, Swedish, Tagalog, and Turkish. The updates apply to both batch and…

AI Voice Agents vs. Human Agents: Metrics Compared

Deepgram argues that legacy contact center metrics like AHT, containment, FCR, CSAT, and QA scorecards were designed for human agents and fail to measure AI voice agents accurately. The post proposes …

Conversation Context for Voice Agents: What It Is and Why It Matters for Accuracy

Deepgram demonstrates that integrating conversation context into speech-to-text (STT) systems significantly reduces word error rates (WER) by up to 24% compared to baseline models like wav2vec2.0 and …

How to Evaluate Voice Quality Claims Across Text-to-Speech Providers

The article critiques vendor-controlled TTS benchmarks, showing how demo conditions, text selection, and methodology bias results. It proposes a repeatable evaluation protocol using ITU-T standards, W…

The real-time voice AI stack for agents: An architecture guide

Deepgram’s guide explains how a real-time voice AI stack must complete each turn within ~1 second to feel natural, with strict latency budgets per layer (transport, STT, LLM, TTS). It compares provide…

Deepgram Self-Hosted August 2026 release adds Flux TTS GA and expressivity controls

Deepgram’s August 26, 2026 self-hosted release (260826) introduces general availability for Flux TTS, new expressivity controls to adjust voice tone, and improved currency, date, and Japanese punctuat…

Pricing change detected for Deepgram

Pricing updated for Deepgram: - Developers & Startups: Custom pricing → $0.0077/mo - Developers & Startups: new promotion — Limited-time promotional rates on streaming - Developers & Startups: added a…

Voice Agents Meet the Phone Line: What It Really Takes to Ship in Production

Deepgram and Twilio announced a joint technical webinar on August 19, 2026, focused on deploying voice agents on phone lines. The session highlights challenges like latency, codecs, voicemail, and com…

Deepgram CLI 0.3.0 adds Flux TTS support, fixes Flux STT streaming, and enforces exit codes

Deepgram released CLI 0.3.0 with full Flux TTS support via `dg speak`, fixes for Flux STT streaming in `dg listen`, and a breaking change enforcing exit codes (0=success, 1=error, 2=user interrupt). T…

Pricing updated for Deepgram: - Pay-As-You-Go: promotion changed — Free $200 Credit then pay-as-you-go - Developers & Startups: $4000/yr → Custom pricing - Developers & Startups: allowance changed fro

Voice AI Infrastructure Leader Deepgram to Establish APAC Headquarters in Singapore Following Strategic Investment from EDBI

Deepgram will open its Asia-Pacific headquarters in Singapore, appointing Sriram Ved as VP and GM for the region. The move follows a 96% YoY increase in APAC API requests and a $130M Series C extensio…

Deepgram Flux TTS for Developers: What Ships at General Availability

Deepgram’s Flux TTS API reached general availability on a new /v2/speak endpoint, replacing the older /v1/speak. The update introduces turn-based streaming for real-time voice agents, batch processing…

Deepgram Nova-3 adds Afrikaans and Georgian with improved models for 7 languages

Deepgram expanded Nova-3 with two new languages (Afrikaans and Georgian) and improved monolingual models for seven existing languages (Russian, Hungarian, Macedonian, Slovak, Slovenian, and Urdu). The…

Why TTS Latency Grows With Response Length and the Architecture That Stops It

The article explains how TTS latency scales with response length due to autoregressive generation dependencies, contrasting it with parallel/non-autoregressive models. It highlights that streaming run…

Deepgram adds Numerals support for Bulgarian, Cantonese, Malay, and Korean

Deepgram expanded its Numerals feature to support four new languages—Bulgarian, Chinese (Cantonese, Traditional), Malay, and Korean—using monolingual models. The feature converts spoken numbers into d…

Pricing updated for Deepgram: - Pay-As-You-Go: promotion changed — $200 free credit - New tier: Developers & Startups ($4000/yr) - New tier: Growing Applications ($4000/yr) - Removed: Custom tier

Introducing Flux STT: Conversational Speech Recognition to Solve the Biggest Problem in Voice Agents-Interruptions

Deepgram introduced Flux STT, the first speech recognition model designed specifically for conversational voice AI, not just transcription. Unlike traditional ASR systems, Flux STT integrates turn det…

Introducing Flux TTS: Conversation-Native Text to Speech for Real Time Voice Agents

Deepgram introduced Flux TTS, the first conversation-native text-to-speech model designed for real-time voice agents, addressing failures of traditional TTS in live interactions. The model maintains c…

Deepgram Flux TTS adds precise interruption handling for voice agents

Deepgram’s Flux TTS now supports precise interruption handling via a new `Interrupt` message that stops audio playback locally and sends a `playback_offset` to the server. The server responds with `Sp…

Deepgram expands Flux TTS voice catalog with 25+ new voices and accents

Deepgram updated its Flux TTS voice catalog, introducing a structured model string format (`flux-{voice}-{language}`) and expanding the roster to over 25 voices across American, British, Irish, Austra…

Text-to-Speech Comes of Age: Deepgram Launches Conversation-Native Speech

Deepgram released Flux TTS, a conversation-native text-to-speech model designed for enterprise voice agents, joining its existing Flux STT within the Voice AI platform. The model maintains context, ad…

Deepgram Self-Hosted August 2026 release adds Flux TTS Early Access and Nepali/Punjabi support

Deepgram’s August 12, 2026 self-hosted release introduces Flux TTS in Early Access for self-hosted deployments via a new /v2/speak endpoint, requiring dedicated GPU resources. It also adds Nova-3 supp…

Deepgram Nova-3 adds Armenian and improves Tamil, Indonesian, and Belarusian models

Deepgram introduced Armenian as a new supported language for its Nova-3 model and improved monolingual models for Tamil, Indonesian, and Belarusian. The updates enhance transcription quality for both …

A Developer's Guide to Fixing Common TTS Pronunciation Errors

Deepgram published a technical guide categorizing TTS pronunciation errors into heteronyms, domain terms, alphanumeric sequences, and text normalization failures. It details SSML phoneme tags, PLS lex…

Deepgram releases improved Nova-3 models for 9 languages

Deepgram launched enhanced Nova-3 monolingual models for nine languages, improving transcription accuracy for both batch and streaming workloads. Updates cover Bulgarian, Bengali, Bosnian, Belarusian,…

How Moveo Benchmarks Multilingual Voice AI with Deepgram for Real Contact Center Calls

Moveo, a provider of contact center voice AI agents, has integrated Deepgram’s streaming speech-to-text (STT) models into its multilingual voice agents, prioritizing accuracy, end-of-turn detection, l…

Voice AI APIs for CRM integration: building the pipeline from call audio to customer data

Deepgram’s guide breaks the CRM voice AI pipeline into four stages—capture, transcription, structuring, and write-back—highlighting how accuracy and architecture decisions determine CRM data quality. …

Deepgram Nova-3 adds Punjabi and Nepali language support

Deepgram expanded its Nova-3 monolingual model to support Punjabi and Nepali, adding language codes `pa`, `pa-IN`, and `ne`. Both batch and streaming models now include these languages, accessible by …

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

Prem AI Brings Sovereign Voice AI to Regulated Enterprises with Deepgram Nova-3

Prem AI, a sovereign AI infrastructure provider, integrated Deepgram's Nova-3 speech-to-text model into its private, self-hosted ecosystem to enable real-time, multilingual voice capabilities for regu…

2X Solutions AI Turns Real-Time Transcription into a Compliance Control for Regulated Voice Outreach with Deepgram

2X Solutions AI adopted Deepgram’s Nova-2 speech-to-text model on-premise to power real-time transcription for compliance-sensitive voice outreach, processing over 7.3 million calls. The integration e…

Pricing updated for Deepgram: - New tier: Developers & Startups ($4000/yr) - New tier: Growing Applications ($4000/yr) - Removed: Custom tier

Deepgram releases improved Nova-3 models for 11 new languages

Deepgram launched enhanced Nova-3 monolingual models for 11 languages, including Croatian, Hungarian, Lithuanian, Korean, Japanese, and others, improving transcription accuracy for both batch and stre…

Voice Agent Orchestration Layer: Enterprise Unbundling Guide

Deepgram’s product marketing manager published a guide arguing that all-in-one voice agent platforms fail at scale due to pricing ceilings, vendor lock-in, and compliance gaps. The piece promotes modu…