Deepgram STT/Voice Models Native on Together AI

Post LinkedIn

🤝Read original on Together AI Blog

#stt #tts #voice-agents #inferencedeepgram-stt/ttsdeepgram together-ai

💡Prod Deepgram STT/TTS native on Together AI—scale real-time voice agents effortlessly.

⚡ 30-Second TL;DR

What Changed

Deepgram STT and TTS models now natively hosted on Together AI

Why It Matters

This lowers barriers for developers building voice AI by integrating Deepgram's accurate models into Together AI's scalable inference. It streamlines multi-modal agent development without vendor switching.

What To Do Next

Test Deepgram STT/TTS on Together AI Dedicated Model Inference for your voice agent prototype.

Who should care:Developers & AI Engineers

Key Points

•Deepgram STT and TTS models now natively hosted on Together AI
•Production-ready for real-world deployment
•Accessible via Dedicated Model Inference
•Optimized for real-time voice agents

🧠 Deep Insight

AI-generated analysis for this event — not the original article.

🔑 Enhanced Key Takeaways

•The integration leverages Together AI's high-throughput inference infrastructure to reduce latency for Deepgram's Nova-2 and Aura models, specifically targeting sub-200ms round-trip times for voice AI applications.
•This partnership allows developers to consolidate their AI stack by running both LLM reasoning and voice processing (STT/TTS) within the same VPC or infrastructure environment, minimizing data egress costs.
•The deployment utilizes Together AI's Dedicated Model Inference (DMI) to provide guaranteed compute resources, ensuring consistent performance for high-volume enterprise voice agents compared to shared-instance alternatives.

📊 Competitor Analysis▸ Show

Feature	Deepgram on Together AI	Groq (Whisper/TTS)	AWS Transcribe/Polly
Latency	Ultra-low (Optimized)	Ultra-low (LPU-based)	Moderate/High
Model Variety	Nova-2 (STT), Aura (TTS)	Whisper, Custom	Broad, General Purpose
Deployment	Dedicated Inference	Serverless/API	Managed Service
Pricing Model	Dedicated Instance	Per-token/Usage	Per-minute/Usage

🛠️ Technical Deep Dive

Architecture: Integration utilizes Deepgram's proprietary Nova-2 (transformer-based ASR) and Aura (fast-streaming TTS) architectures optimized for Together AI's GPU clusters.
Inference Stack: Leverages Together AI's custom inference engine, which supports optimized CUDA kernels for faster token generation in TTS and reduced decoding latency for STT.
Integration: Models are containerized and deployed via Together AI's DMI, allowing for direct API access that mirrors standard OpenAI-compatible endpoints for easier developer migration.
Performance: Designed to support streaming audio input/output, facilitating full-duplex communication necessary for conversational AI agents.

🔮 Future ImplicationsAI analysis grounded in cited sources

Voice agent development costs will decrease by 15-20% for high-volume users.

Consolidating STT, TTS, and LLM inference on a single infrastructure provider eliminates multi-vendor data transfer fees and optimizes compute utilization.

Real-time voice latency will become the primary competitive differentiator for AI agent platforms.

As model intelligence reaches parity, the ability to minimize conversational lag through integrated infrastructure will dictate user adoption in customer service and enterprise automation.

⏳ Timeline

2023-11

Deepgram releases Nova-2, their flagship speech-to-text model.

2024-05

Deepgram launches Aura, their text-to-speech model focused on conversational speed.

2025-02

Together AI expands Dedicated Model Inference (DMI) to support third-party model hosting.

2026-04

Deepgram STT/TTS models become natively available on Together AI infrastructure.

🤝Read original article on Together AI Blog

📰

Weekly AI Recap

Read this week's curated digest of top AI events →

👉Related Updates

Same topic

Explore #stt

Same product