iiinova — where your digital self keeps evolving

Deepgram

@deepgram.com

Power your apps with world-class speech and domain-specific language models (DSLMs). Effortlessly accurate. Blazing fast. Enterprise scale. Unbeatable pricing.

Posts Replies Media

Deepgram @deepgram.com · 17/09/2026

Introducing Nova-3 Pharma, the first speech-to-text model purpose-built for the pharmaceutical industry. It’s optimized for drug names, pharmaceutical terminology, and medical vocabulary. 🧵👇

111

Deepgram @deepgram.com · 11/09/2026

Speak '26 registration is officially open! 🔥 Voice AI is on the verge of its biggest milestone yet. On Oct 29, the people building it are getting together in SF to talk about what it takes to get there. Register: deepgram.com/speak

011

Deepgram @deepgram.com · 28/08/2026

Clear your calendar for October 29th. Deepgram Speak ’26 is landing in San Francisco. Save the date → deepgram.com/speak

010

Deepgram @deepgram.com · 12/08/2026

Today, we're launching Flux TTS: expressive voices built for live conversation. Most TTS handles one line at a time. Flux TTS carries context across the whole conversation. And it's a conversation you'll actually want to have. Hear how Flux meets each moment. talk.deepgram.com

020

Deepgram @deepgram.com · 10/08/2026

just got the text new voices are entering the villa and they just might turn some heads~ august 12th stay tuned!!! 👀

011

Deepgram @deepgram.com · 21/07/2026

Your party deserves a historian. 📜 Jabari Barton built CritScribbler so D&D players never lose the plot. Session audio goes in, searchable notes and campaign lore come out, built by Deepgram. We love when our community builds the tool they wished existed ⚔️🎲

011

Deepgram @deepgram.com · 29/06/2026

First Developer Spotlight is live! And we're opening with one of our favorites. CATT: a real-time transcription and translation tool for VTubers from Japan and Korea streaming to Western and SEA audiences, built with Deepgram.

140

Deepgram @deepgram.com · 17/06/2026

The Deepgram Australia endpoint is now generally available. api.au.deepgram.com runs in Sydney, with your audio, transcripts, and speech output processed and stored entirely within Australia. Same API, same models. Moving over is a base URL change. Blog → deepgram.com/learn/deepgr...

000

Deepgram @deepgram.com · 10/06/2026

Introducing Batch Diarization V2, a major upgrade to speaker labeling for pre-recorded audio. Understanding what was said is only part of the story. You also need to know who said it.

Deepgram Batch Diarization V2 Improved speaker attribution accuracy

210

Deepgram @deepgram.com · 07/05/2026

Conversational latency matters just as much as transcription accuracy. Across aggregate end-of-turn benchmarks, Flux Multilingual delivers: ⚡ Highest aggregate EoT F1 across all supported languages ⚡ Up to 3x lower latency than competing real-time EoT systems

Scatter plot titled "EoT Competitive Benchmark—AGGREGATE (all languages pooled)" showing Turn Precision (y-axis, ranging from 0.55 to 0.90) versus Median EoT Latency P50 in milliseconds (x-axis, ranging from 0 to 2000). Flux Multilingual (teal circles) clusters in the upper-left, achieving high turn precision at low latency. Assembly Universal 3 Pro (pink cross) sits near 750ms latency with ~0.72 turn precision. ElevenLabs Scribe V2 silence sweep (filled blue squares) and VAD sweep (hollow blue squares) appear in the lower-right, with high latency around 1500–2000ms and turn precision between 0.60 and 0.70.

Scatter plot titled "EoT Competitive Benchmark—AGGREGATE (all languages pooled)" showing Turn F1 score (y-axis, ranging from 0.70 to 0.86) versus Median EoT Latency P50 in milliseconds (x-axis, ranging from 0 to 2000). Flux Multilingual (teal circles) clusters in the upper-left quadrant, achieving high F1 scores between 0.78 and 0.85 at latencies under 500ms. Assembly Universal 3 Pro (pink cross) sits near 750ms with an F1 of ~0.82. ElevenLabs Scribe V2 silence sweep (filled blue squares) and VAD sweep (hollow blue squares) appear in the lower-right at latencies around 900–2000ms, with F1 scores between 0.70 and 0.81.

110

Deepgram @deepgram.com · 07/05/2026

Bar chart titled "Portuguese: Word Error Rate (WER)" with subtitle "Real-world production data (lower is better)". The y-axis shows Word Error Rate as a percentage, ranging from 0 to 40. Eight models are compared: Deepgram Flux Multilingual (13.16%, teal), Azure Auto (14.20%, purple), Soniox STT RT v4 (14.91%, gold), Speechmatics Enhanced (15.76%, blue), AssemblyAI Pro v3 (16.94%, yellow), ElevenLabs Scribe v2 Realtime (18.02%, pink), Google Chirp 3 (19.53%, grey), and OpenAI GPT-4o Transcribe (21.72%, orange). Deepgram Flux Multilingual achieves the lowest WER.

Bar chart titled "Hindi: Word Error Rate (WER)" with subtitle "Real-world production data (lower is better)". The y-axis shows Word Error Rate as a percentage, ranging from 0 to 70. Six models are compared (AssemblyAI data not available): Deepgram Flux Multilingual (17.18%, teal), Soniox STT RT v4 (28.16%, gold), Speechmatics Enhanced (32.66%, blue), Azure Auto (33.58%, purple), Google Chirp 3 (34.18%, grey), OpenAI GPT-4o Transcribe (53.42%, orange), and ElevenLabs Scribe v2 Realtime (61.83%, pink). A footnote notes that AssemblyAI data does not exist. Deepgram Flux Multilingual achieves the lowest WER by a wide margin.

100

Deepgram @deepgram.com · 07/05/2026

We benchmarked Flux Multilingual on real-world production audio across all supported languages using each vendor’s default streaming configuration. Result: 🏆 Best-in-class WER across the majority of supported languages, including: 🇺🇸 English 🇪🇸 Spanish 🇩🇪 German 🇫🇷 French 🇵🇹 Portuguese 🇮🇳 Hindi

Bar chart titled "English: Word Error Rate (WER)" with subtitle "Real-world production data (lower is better)". The y-axis shows Word Error Rate as a percentage, ranging from 0 to 40. Eight models are compared: Deepgram Flux Multilingual (9.94%, teal), AssemblyAI Pro v3 (11.60%, yellow), Soniox STT RT v4 (11.87%, gold), Speechmatics Enhanced (11.93%, blue), Azure Auto (13.05%, purple), Google Chirp 3 (15.65%, grey), ElevenLabs Scribe v2 Realtime (16.49%, pink), and OpenAI GPT-4o Transcribe (30.73%, orange). Deepgram Flux Multilingual achieves the lowest WER.

Bar chart titled "Spanish: Word Error Rate (WER)" with subtitle "Real-world production data (lower is better)". The y-axis shows Word Error Rate as a percentage, ranging from 0 to 30. Eight models are compared: Deepgram Flux Multilingual (11.29%, teal), Speechmatics Enhanced (11.32%, blue), Soniox STT RT v4 (12.07%, gold), AssemblyAI Pro v3 (13.80%, yellow), Azure Auto (14.81%, purple), ElevenLabs Scribe v2 Realtime (15.71%, pink), Google Chirp 3 (16.41%, grey), and OpenAI GPT-4o Transcribe (23.74%, orange). Deepgram Flux Multilingual achieves the lowest WER.

Bar chart titled "German: Word Error Rate (WER)" with subtitle "Real-world production data (lower is better)". The y-axis shows Word Error Rate as a percentage, ranging from 0 to 40. Eight models are compared: Deepgram Flux Multilingual (10.66%, teal), Soniox STT RT v4 (11.85%, gold), AssemblyAI Pro v3 (12.49%, yellow), Speechmatics Enhanced (12.58%, blue), Google Chirp 3 (15.05%, grey), Azure Auto (15.42%, purple), ElevenLabs Scribe v2 Realtime (19.63%, pink), and OpenAI GPT-4o Transcribe (26.48%, orange). Deepgram Flux Multilingual achieves the lowest WER.

Bar chart titled "French: Word Error Rate (WER)" with subtitle "Real-world production data (lower is better)". The y-axis shows Word Error Rate as a percentage, ranging from 0 to 40. Eight models are compared: Deepgram Flux Multilingual (13.62%, teal), Azure Auto (14.81%, purple), Speechmatics Enhanced (15.86%, blue), Soniox STT RT v4 (16.03%, gold), Google Chirp 3 (16.23%, grey), AssemblyAI Pro v3 (16.74%, yellow), ElevenLabs Scribe v2 Realtime (19.06%, pink), and OpenAI GPT-4o Transcribe (23.64%, orange). Deepgram Flux Multilingual achieves the lowest WER.

100

Deepgram @deepgram.com · 05/05/2026

The dg CLI, MCP server, and deepgram/skills repo shipped in April. Together they make Deepgram a first-class citizen in Claude Code, Cursor, Windsurf, Codex, and Aider. What you can actually build with them: deepgram.com/learn/agenti... CLI cli.deepgram.com Skills github.com/deepgram/ski...

120

Deepgram @deepgram.com · 03/03/2026

Flux now supports on-the-fly configuration. Update keyterms and turn detection thresholds mid-stream on your existing WebSocket connection, no reconnecting required. One connection that shifts context as the conversation evolves. Blog → deepgram.com/learn/flux-o...

021