Voice AI

Deepgram Speech-to-Text: The Complete 2026 Guide

Everything a builder needs to know about Deepgram's STT models — Nova-3, Flux, pricing structure, accuracy tradeoffs, and how to actually integrate it into a voice AI product.

Written by

Akash Maurya

July 15, 2026

13 min read

Deepgram Speech-to-Text: The Complete 2026 Guide

Voice AI

Read article

Deepgram is the default STT choice for most voice AI builders in 2026, and for good reason — but 'default choice' doesn't mean 'the right choice for every use case.' This guide covers when Deepgram is genuinely the best option, when a competitor makes more sense, and how to avoid the pricing surprises that catch teams off guard once diarization, redaction, and multilingual add-ons stack up.

This guide breaks down Deepgram's model lineup, real per-minute pricing (not just the headline number), streaming vs batch tradeoffs, and the integration patterns that actually matter when you're building a production voice agent rather than a transcription demo.

Nova-3 vs Flux: Which Model Should You Use

Nova-3 is Deepgram's flagship general-purpose model — strong accuracy across 50+ languages, available in both streaming and pre-recorded modes. Flux is a newer, more specialized model built specifically around the hardest problem in live voice agents: knowing when someone has actually finished speaking versus just paused. If you're transcribing recordings after the fact, Nova-3 is the right (and cheaper) choice. If you're building a live conversational agent where turn-taking quality directly affects how natural the conversation feels, Flux is worth the roughly 40-70% price premium.

Model	Best For	Approx. Streaming Rate	Languages
Nova-3 Monolingual	General transcription, single language	$0.0077/min	English-optimized
Nova-3 Multilingual	Mixed-language or unpredictable-language audio	Higher than monolingual	50+ languages
Flux English	Live voice agents needing precise turn detection	$0.0065/min	English
Flux Multilingual	Live multilingual voice agents	$0.0078/min	Multiple

Streaming vs Batch: Pick Based on Latency Need, Not Habit

A surprising number of teams default to streaming transcription even for workloads that don't need it — like overnight call-recording analysis — and pay roughly 1.5-2x more per minute for latency they're not using. If your use case can tolerate even a few minutes of delay (QA review, analytics, compliance archiving), batch/pre-recorded transcription is meaningfully cheaper and often more accurate, since the model can use full audio context rather than processing in small real-time chunks.

Pro Tip

Rule of thumb: if a human is waiting on the transcript in real time (a live voice agent, a live captioning feature), use streaming. If the transcript is consumed later (QA, analytics, search), use batch and save 30-50% on that line item.

The Real Cost of Deepgram — Beyond the Headline Rate

The advertised $0.0077/min (Pay-As-You-Go, Nova-3 streaming) or $0.0043/min (pre-recorded) is only the base transcription cost. Diarization (speaker identification), PII redaction, and multichannel audio (common in call-center recordings, where each side of the call is a separate channel) all add cost on top.

Add-on	Impact on Cost
Multichannel audio (e.g., stereo call recording)	Cost multiplies by number of channels
Speaker diarization	Additional per-minute charge, not included in base rate
PII redaction	Additional per-minute charge
Audio Intelligence (summarization, sentiment)	Separate per-token pricing (~$0.0003-0.0006/1K tokens)

Deepgram vs Competitors

Provider	Approx. Rate (batch)	Strength	Weakness
Deepgram Nova-3	$0.0043/min	Best real-time latency, per-second billing	Not the cheapest raw batch rate
AssemblyAI	$0.0025/min	Cheapest batch rate, strong on noisy audio, built-in summarization	Higher real-time pricing
OpenAI Whisper API	$0.006/min	Open-source option available, wide language coverage	Batch-only, no streaming or diarization
Google Cloud STT	$0.012-0.02/min	Deep GCP ecosystem integration	More expensive, rounding-based billing

Integration Best Practices

Use streaming only where real-time response is actually required — default to batch otherwise
Budget for diarization and redaction separately; don't assume they're bundled
Use the $200 free credit to run your actual production audio (not demo clips) through the model before committing budget
Move to the Growth plan only once your usage consistently exceeds the $4,000/year minimum spend — it doesn't pay off below that
For voice agents specifically, evaluate Flux against Nova-3 on your real conversational audio, not just benchmark clips — turn-detection quality is highly use-case dependent
Watch for multichannel audio silently doubling your bill on stereo call recordings

Common Mistakes

Defaulting to streaming for batch-appropriate workloads and overpaying by 50%+
Not accounting for diarization/redaction add-ons when budgeting, then getting an unexpectedly high first invoice
Comparing Deepgram's per-minute rate directly to AssemblyAI's per-hour rate without converting to the same unit
Skipping evaluation on real production audio and trusting benchmark accuracy numbers, which are self-published and don't reflect your specific accents, noise levels, or domain vocabulary
Not using keyterm prompting for domain-specific vocabulary (medical terms, product names), which meaningfully improves accuracy for a small config change

FAQs

Does Deepgram charge for silence?

No — billing is based on actual audio duration processed, and Deepgram bills per second rather than rounding up, which reduces cost on short clips compared to competitors that round to 15-second increments.

Is Deepgram good for Indian languages?

Deepgram's multilingual coverage is broad but not India-specialized. For Hindi, Hinglish, and other Indian-language-heavy use cases, a purpose-built model like Sarvam AI's Saaras often performs better on code-switching and regional accents.

What's the real cost for a voice agent handling 10,000 minutes/month?

Using Flux English streaming at $0.0065/min, base STT cost alone is roughly $65/month — before diarization or any add-ons, and before your LLM, TTS, and telephony costs.

Problem

Speech-to-text is the first link in any voice AI pipeline, and its latency and accuracy set a ceiling for everything downstream — a slow or inaccurate STT model makes even the best LLM and TTS feel broken. At the same time, STT pricing pages are notoriously confusing, mixing per-minute, per-hour, and per-second billing with add-on fees that aren't in the headline rate.

Solution

Deepgram offers purpose-built models for different needs — Nova-3 for general-purpose high-accuracy transcription, and Flux specifically engineered for the turn-taking problem in live voice agents — with per-second billing that avoids the rounding overhead of competitors, and a genuinely generous free tier for evaluation before committing budget.

Key Features

Nova-3: high-accuracy general transcription, 50+ languages
Flux: turn-detection-aware model built for voice agent conversations
True per-second billing (no 15-second rounding like some competitors)
Real-time streaming and batch/pre-recorded modes
Built-in diarization, redaction, and keyterm prompting (billed separately)
$200 free credit with no expiration on the Pay-As-You-Go tier

Results

Sub-300ms median latency for streaming transcription
Nova-3 batch pricing roughly 2-3x cheaper than Google Cloud STT, 3-5x cheaper than AWS Transcribe
Per-second billing typically saves 5-15% versus rounded-billing competitors on short-call workloads

Technologies Used

DeepgramLiveKitPythonNode.js

Need an AI Solution for Your Business?

I build modern AI Voice Agents, SaaS platforms, automation systems, and full-stack applications that help businesses automate operations and improve customer experiences.

Start a Project View Projects