Deepgram
Listed at https://deepgram.com
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | STT: $0.0043-0.012/min; TTS: $0.015-0.030/1K chars | per million tokens |
| Output | N/A | per million tokens |
🤖 Supported Models (8)
✨ Pros
- ✓Nova-3 STT: 54% WER reduction, industry-leading accuracy
- ✓50+ STT languages including Mandarin and Cantonese
- ✓Sub-300ms STT latency, sub-200ms TTS latency
- ✓$200 free credits, no credit card required
- ✓Full voice stack: STT + TTS + Voice Agent API
- ✓SOC 2/HIPAA/GDPR/PCI compliant
- ✓Self-hosted deployment option available
- ✓Startup program: up to $100K in credits
⚠️ Cons
- ×TTS only supports 7 languages (no Chinese)
- ×No mainland China data center — proxy required
- ×No affiliate program available
- ×Add-on features (custom vocabulary, speaker diarization) cost extra
- ×Billed per minute/character — not directly comparable to LLM token pricing
🎯 Best For
High-accuracy STT applications (meeting transcription, call center analytics, captioning); low-latency TTS for voice assistants; multilingual audio processing; self-hosted voice AI deployments
💰 Pricing & Plans
| Plan / Model | Billing Unit | Pay As You Go | Growth (pre-paid) | Notes |
|---|---|---|---|---|
| Pay As You Go | $200 free credit + usage | $0 upfront | — | No minimums, no credit card required, no expiration |
| Growth | Pre-paid annual credits | — | from $4K/year (save ~20%) | Higher concurrency limits; ideal for growing apps |
| Enterprise | Custom | Contact Sales | Contact Sales | Volume commitments, custom models, dedicated/self-hosted, SLAs |
| STT — Nova-3 Monolingual (Pre-Recorded) | $/min | $0.0043 | $0.0036 | Recommended default; highest accuracy for single-language audio |
| STT — Nova-3 Multilingual (Pre-Recorded) | $/min | $0.0052 | $0.0043 | Automatic language detection; 45+ languages |
| STT — Flux English (Streaming) | $/min | $0.0065 (promo) | $0.0057 | Conversational STT with turn detection; promo price |
| STT — Flux Multilingual (Streaming) | $/min | $0.0078 | $0.0068 | Multi-language conversational STT, single-stream |
| STT — Whisper Cloud (Pre-Recorded) | $/min | $0.0048 | $0.0048 | OpenAI Whisper hosted; Deepgram concurrency limit 5 |
| TTS — Aura-1 | $/1K chars | $0.0150 | $0.0135 | Legacy model, broad voice catalog |
| TTS — Aura-2 | $/1K chars | $0.030 | $0.027 | Latest natural-sounding voices, low-latency |
| TTS — Flux TTS | $/1K chars | $0.0450 | — | Conversational TTS with sub-200ms latency |
| Voice Agent — Standard | $/min (WebSocket) | $0.075 | $0.068 | Default tier; Deepgram-managed LLM + TTS |
| Voice Agent — Standard (BYO TTS) | $/min | $0.065 | $0.051 | Bring your own TTS provider |
| Voice Agent — Advanced | $/min | $0.163 | $0.146 | Higher quality, advanced turn-taking |
| Audio Intelligence — Summarization | $/1K tokens | $0.0003 in / $0.0006 out | $0.00024 in / $0.00048 out | LLM-powered audio summary |
🔧 API & Developer Experience
- •Protocol: REST + WebSocket dual transport — REST for batch STT/TTS generation, WebSocket for low-latency streaming and Voice Agent real-time conversation; sub-300ms first-byte on streaming STT
- •Official SDKs: Python (`pip install deepgram-sdk`), Node.js (`npm install @deepgram/sdk`), Go, .NET, Rust — official SDKs maintained by Deepgram; well-documented at developers.deepgram.com
- •API Playground: Hosted API Playground at playground.deepgram.com — try Nova-3 / Flux / Aura-2 models with audio upload or mic input, no code required
- •Audio Intelligence add-ons: Sentiment Analysis, Topic Detection, Intent Recognition, Summarization, Smart Formatting, Speaker Diarization, Entity Detection, Redaction — modular $/min or $/1K-token add-ons to STT calls
- •Self-hosted deployment: In-VPC or fully self-hosted option for Enterprise tier — same Nova-3 model weights run on customer infrastructure for regulated industries (HIPAA BAA, data residency)
- •MCP server & AI integrations: Native integration with IBM watsonx Orchestrate, Five9, Genesys, Twilio, Vonage, Amazon Connect, Kore.ai, AudioCodes; MCP-friendly structured JSON output
- •Free credit & billing model: $200 free credit at signup with no credit card; usage-based per-minute (STT, Voice Agent) or per-1K characters (TTS); Audio Intelligence billed per 1K input/output tokens
- •Compliance & data residency: SOC 2 Type 1 & Type 2 certified, HIPAA-compliant with BAA available for Enterprise; data residency options in US, EU, and Australia regions
🎙️ Speech Accuracy & Real-Time Voice Agent Infrastructure
Deepgram's defining edge is end-to-end voice infrastructure: rather than stitching STT, TTS, and an agent orchestrator from separate vendors, Deepgram ships Nova-3 (best-in-class WER), Aura-2 (natural-sounding TTS), and a managed WebSocket Voice Agent API that wires turn detection, interruption handling, and tool calling into a single low-latency pipeline. Independent benchmarks position Nova-3 as 23-53% more accurate than Amazon Transcribe, Google Cloud STT, and Azure, with 5-40x latency advantage on streaming. The architecture matters for production voice agents because most latency in conversational AI comes from orchestration glue between STT and LLM, and the same model advertises sub-300ms STT first-byte with sub-200ms TTS on Flux. The trade-off is API surface complexity: STT add-ons (diarization, redaction, summarization, intent recognition) are billed separately and can stack quickly in call-center or compliance-heavy workloads, and TTS still supports just seven (no Chinese voice yet) despite 45+ STT languages. For teams that need a single vendor across the voice stack, Deepgram is the safest default; for those who want best-of-breed per layer or Chinese TTS, mixing providers may still be preferable.
🌐 China Access & Latency
Deepgram is a US-headquartered voice AI platform with public API endpoints in the US, EU, and Australia — no mainland China regional edge. Production deployments from mainland China require a stable overseas proxy or an aggregator fronting the Deepgram API; direct connections typically time out or face elevated latency. Real-world latency from China is bounded by the cross-border round trip rather than Deepgram's own sub-300ms streaming capability. Deepgram's WebSocket endpoint is also sensitive to connection drops caused by Chinese carrier middleboxes, which can affect long-running Voice Agent sessions. There is no China-specific pricing, billing, or data-residency option. STT supports Mandarin and Cantonese among 45+ languages, but TTS still lacks Chinese voices. For teams targeting Chinese end users or low-latency interactive voice agents, a domestic STT/TTS vendor (iFLYTEK, Alibaba Voice, Tencent Cloud ASR) plus a self-hosted Voice Agent glue layer is the latency-safe default; Deepgram remains attractive for English/multilingual workloads, US/EU data residency, and HIPAA-regulated healthcare use cases.