Deepgram API 2026: Nova-3 STT & Aura-2 TTS
Deepgram is one of the most respected names in AI speech recognition and synthesis, offering industry-leading speech-to-text (STT) with its Nova-3 model and text-to-speech (TTS) with Aura-2. With sub-300ms STT latency, 50+ languages, and a developer-first API design, Deepgram powers transcription, voice assistants, call center analytics, and accessibility tools for thousands of companies including NASA, Spotify, and The Washington Post. This review covers the full Deepgram API platform, Nova-3 accuracy, Aura-2 voice quality, pricing breakdown, competitive positioning, and developer experience.
TL;DR: Deepgram offers STT from $0.0043/min (Nova-3 streaming) and TTS from $0.015/1K chars (Aura-1). Nova-3 delivers 54% WER reduction over previous models — best-in-class accuracy for English STT. Sub-300ms real-time latency, 50+ languages for STT, $200 free credits on signup. TTS supports 7 languages (no Chinese). No China data center. SOC 2/HIPAA/GDPR compliant. For China-direct multi-provider voice API routing, consider FreeModel.
Introduction: Deepgram — The Developer-First Voice AI Platform
Deepgram was founded in 2015 with a mission to make voice data accessible and actionable through AI. Unlike many voice API providers that started as consumer products and added APIs later, Deepgram was built API-first from day one. Their core differentiator is end-to-end deep learning: instead of stitching together traditional speech recognition pipelines (acoustic model + language model + decoder), Deepgram trains a single neural network end-to-end on raw audio.
This architectural choice pays off in accuracy and latency. The latest Nova-3 model achieves a 54% word error rate (WER) reduction over the previous generation, making it one of the most accurate speech recognition APIs available. Combined with sub-300ms streaming latency, Deepgram is a strong choice for real-time transcription, voice assistants, and any application where both speed and accuracy matter.
Deepgram serves a wide range of industries: media and entertainment (captions, transcription), contact centers (real-time agent assist, analytics), healthcare (medical transcription, HIPAA-compliant), legal (court reporting, deposition transcription), and accessibility (captioning for the deaf and hard of hearing).
What Deepgram Actually Offers
Deepgram provides a comprehensive voice AI API platform centered on speech-to-text and text-to-speech:
Speech-to-Text (STT) — Nova-3 & Flux
Deepgram's flagship product is its STT API. Three model families cover different use cases:
| Model | Type | Latency | Languages | Pricing (per min) |
|---|---|---|---|---|
| Nova-3 Monolingual | Best accuracy | under 300ms streaming | English only | $0.0043 streaming / $0.0048 pre-recorded |
| Nova-3 Multilingual | Multi-language | under 300ms streaming | 50+ languages | $0.0052 streaming / $0.0058 pre-recorded |
| Flux | Conversational | under 300ms | 50+ languages | $0.0065 streaming / $0.0077 pre-recorded |
| Whisper | Open model | Higher latency | 99+ languages | $0.0048/min |
Nova-3 Monolingual (English) has the best accuracy, with Deepgram reporting 54% WER reduction over Nova-2. Nova-3 Multilingual supports 50+ languages including Mandarin Chinese, Cantonese, Japanese, Korean, Spanish, French, German, and more. Flux is optimized for conversational speech with native turn detection — ideal for call center and voice assistant use cases.
Key STT features include: real-time streaming via WebSocket, pre-recorded batch processing, speaker diarization (who said what), custom vocabulary (domain-specific terms), punctuation and formatting, profanity filtering, and language detection.
Text-to-Speech (TTS) — Aura-2 & Aura-1
Deepgram's TTS offering, branded as Aura, provides natural-sounding speech synthesis:
| Model | Latency | Quality | Languages | Pricing (per 1K chars) |
|---|---|---|---|---|
| Aura-2 | Sub-200ms | Best (expressive, natural) | 7 languages | $0.030 |
| Aura-1 | Sub-200ms | High | 7 languages | $0.015 |
Aura-2 offers more expressive, natural-sounding speech with better prosody and emotional range. Both models support SSML for fine-grained control over pronunciation, emphasis, and pacing. Languages include English, Spanish, French, German, Portuguese, Japanese, and Korean. Notably, Chinese (Mandarin or Cantonese) is not supported for TTS.
Voice Agent API
A newer product that combines STT + TTS + LLM into a unified voice agent pipeline. Developers can create conversational voice agents with turn detection, knowledge base integration, and real-time voice-to-voice conversation — all through a single API stream.
Pricing: Pay-as-You-Go with $200 Free Credits
Deepgram uses a straightforward pay-as-you-go pricing model — no confusing credit systems, no tiers:
| Service | Model | Pay-as-You-Go | Growth Plan (~12-20% off) |
|---|---|---|---|
| STT (streaming) | Nova-3 Mono | $0.0043/min | $0.0036/min |
| STT (pre-recorded) | Nova-3 Mono | $0.0048/min | $0.0040/min |
| STT (streaming) | Nova-3 Multi | $0.0052/min | $0.0044/min |
| STT (streaming) | Flux | $0.0065/min | $0.0055/min |
| TTS | Aura-2 | $0.030/1K chars | $0.027/1K chars |
| TTS | Aura-1 | $0.015/1K chars | $0.013/1K chars |
The Growth Plan (~$1,000+/month) offers 12-20% volume discounts. Deepgram also provides a Startup Program with up to $100K in credits for qualifying startups. All new accounts receive $200 in free credits to test the API — no credit card required.
Accuracy: Nova-3 Benchmark — 54% WER Reduction
Nova-3 represents a significant leap in STT accuracy. Deepgram claims a 54% word error rate (WER) reduction over Nova-2 on internal benchmarks. In practical terms, this means:
- Clean audio (studio podcast): WER ~3-5% — nearly human-level transcription
- Conversational (call center): WER ~8-12% — usable for real-time agent assist
- Noisy environments (field recording): WER ~12-18% — significantly better than competitors
The end-to-end deep learning architecture handles background noise, multiple speakers, accents, and domain-specific vocabulary better than traditional pipeline-based approaches.
API Surface and Developer Experience
Deepgram provides a clean REST API for pre-recorded audio and WebSocket API for real-time streaming. SDKs are available for Python, JavaScript/Node.js, Go, .NET, and Java:
import requests
# STT - Pre-recorded audio
url = "https://api.deepgram.com/v1/listen?model=nova-3"
headers = {
"Authorization": "Token YOUR_DEEPGRAM_API_KEY",
"Content-Type": "audio/wav"
}
with open("meeting.wav", "rb") as audio:
response = requests.post(url, headers=headers, data=audio)
result = response.json()
print(result["results"]["channels"][0]["alternatives"][0]["transcript"])# TTS - Text-to-Speech
url = "https://api.deepgram.com/v1/speak?model=aura-2"
headers = {
"Authorization": "Token YOUR_DEEPGRAM_API_KEY",
"Content-Type": "application/json"
}
data = {"text": "Hello from Deepgram Aura-2."}
response = requests.post(url, headers=headers, json=data)
with open("output.mp3", "wb") as f:
f.write(response.content)Developer experience is a strong point: clear documentation with interactive examples, a web-based Playground for testing models without code, and detailed benchmarking data published transparently. Deepgram also provides a self-hosted option for on-premise or air-gapped deployments (SOC 2/HIPAA compliant), which is rare among voice AI API providers.
Pros and Cons
Pros
- ✅ Nova-3 STT accuracy is best-in-class (54% WER reduction)
- ✅ Sub-300ms streaming latency — real-time capable
- ✅ Straightforward pay-as-you-go pricing with $200 free credits
- ✅ 50+ STT languages including Mandarin and Cantonese
- ✅ Developer-friendly API with excellent documentation
- ✅ Self-hosted option for air-gapped/on-premise deployments
- ✅ SOC 2/HIPAA/GDPR/PCI compliant — enterprise-ready
- ✅ Startup program with up to $100K in credits
Cons
- ❌ TTS only supports 7 languages — no Chinese support
- ❌ No mainland China data center — proxy required from China
- ❌ No affiliate/referral program available
- ❌ Add-on features (custom vocabulary, advanced diarization) cost extra
- ❌ Not an LLM API — voice processing only, no text generation
- ❌ Whisper API has concurrency limits and higher latency
Use Case Recommendations
| Use Case | Recommended Model | Why |
|---|---|---|
| Real-time meeting transcription | Nova-3 Monolingual (streaming) | Best accuracy + sub-300ms latency |
| Call center analytics | Nova-3 Multilingual or Flux | Speaker diarization + turn detection |
| Video captioning / subtitles | Nova-3 Multilingual (pre-recorded) | Batch processing, 50+ languages |
| Voice assistants / chatbots | Nova-3 STT + Aura-2 TTS | Full duplex voice pipeline |
| Medical / legal transcription | Nova-3 + custom vocabulary | Domain-specific term accuracy |
| Self-hosted / air-gapped deployment | Deepgram Self-Hosted | On-premise, SOC 2/HIPAA |
| China-direct voice API access | FreeModel aggregator | Multi-provider routing, China direct |
Comparison: Deepgram vs Competitors
| Provider | Best For | STT Accuracy | China Access | STT Pricing (per min) |
|---|---|---|---|---|
| Deepgram | Best STT accuracy, developer-first | Best (Nova-3) | Proxy required | $0.0043-0.0077 |
| ElevenLabs Scribe | TTS-centric voice AI stack | Very good | Proxy required | $0.13/min |
| Azure Speech | Enterprise, 140+ languages | Good | Proxy required | $0.006-0.024 |
| Google Cloud STT | Google ecosystem | Good | Blocked | $0.006-0.024 |
| OpenAI Whisper | Open model, 99+ languages | Good | Proxy required | $0.006/min |
| FreeModel | Multi-provider aggregator, China direct | Varies | Direct | Varies by provider |
FAQ
Q: How accurate is Deepgram Nova-3 for English STT?
A: Nova-3 Monolingual achieves the best-in-class word error rate (WER) in its class, with Deepgram reporting 54% WER reduction over Nova-2. In practical terms, clean audio transcribes at ~3-5% WER (near human-level), conversational speech at ~8-12%, and noisy field recordings at ~12-18%.
Q: Does Deepgram support Chinese languages?
A: Yes for STT — Nova-3 Multilingual and Flux support Mandarin Chinese and Cantonese. No for TTS — Aura-2 and Aura-1 do not include Chinese voices (only English, Spanish, French, German, Portuguese, Japanese, and Korean).
Q: Can I use Deepgram from China?
A: Yes, but a stable proxy or VPN connection is recommended. Deepgram's API endpoints are in the US and EU only — there is no mainland China data center. For teams needing direct China access, consider FreeModel as an aggregator with multi-provider voice API routing.
Q: How does Deepgram pricing compare to Google Cloud STT or Azure Speech?
A: Deepgram is generally more cost-effective for high-volume STT. Nova-3 at $0.0043/min (streaming) is significantly cheaper than Google Cloud ($0.006-0.024/min) and Azure Speech ($0.006-0.024/min) for comparable accuracy. The $200 free credits also make testing accessible without upfront investment.
Q: Does Deepgram offer real-time streaming?
A: Yes. Deepgram's core differentiator is real-time streaming via WebSocket with sub-300ms latency. Both STT (Nova-3, Flux) and TTS (Aura-2, Aura-1) support streaming. The pre-recorded API is available for batch processing.
Q: Is Deepgram SOC 2 and HIPAA compliant?
A: Yes. Deepgram is SOC 2 Type II certified, HIPAA compliant, and GDPR compliant. The self-hosted option allows air-gapped deployments for the strictest security requirements. PCI compliance is also available for payment-related voice processing.
Q: Does Deepgram have an affiliate or referral program?
A: No. Deepgram does not currently offer a commission-based affiliate program. They have a Partner Program (Technology, Channel, Consultancy) for integration and reselling partnerships, but no referral commission model for individual developers or publishers.
Conclusion
Deepgram is the go-to choice for developers who prioritize STT accuracy and latency above all else. Nova-3 Monolingual is the most accurate English speech recognition API available at a competitive price point, and the developer experience is excellent — clean APIs, thorough documentation, and a generous $200 free tier. The self-hosted option is a standout for enterprise customers with strict data residency requirements.
The main limitations are the narrow TTS language support (7 languages, no Chinese) and the lack of a mainland China data center. For teams needing China-direct access or multi-provider voice API routing through a single endpoint, FreeModel at freemodel.dev/invite/FRE-7a3b6220 provides a complementary aggregator solution.
Comparison Table (Final)
| Provider | Best For | STT Starting Price | China Access |
|---|---|---|---|
| Deepgram | Best STT accuracy, low latency | $0.0043/min (streaming) | Proxy required |
| ElevenLabs | Best TTS, full voice AI stack | $0.13/min (STT Scribe) | Proxy required |
| Azure Speech | Enterprise, 140+ languages | $0.006/min | Proxy required |
| Google Cloud STT | Google ecosystem | $0.006/min | Blocked |
| FreeModel | Multi-provider aggregator, China direct | Free credits | Direct |