What Deepgram does
Deepgram offers streaming and prerecorded speech recognition, text-to-speech voices, audio intelligence, and voice-agent orchestration through hosted or enterprise deployments.
Deepgram spans three layers of a voice application. Nova and Flux models transcribe prerecorded or live speech, Aura generates speech from text, and the Voice Agent API coordinates a conversational loop. Add-ons can supply diarization, key-term prompting, redaction, summarization, intent, topics, or sentiment, while enterprise customers can discuss self-hosted deployments.
Pay As You Go currently starts with $200 credit and no minimum. The live table lists model-specific per-minute speech rates, character-based TTS, token-priced intelligence, and conversation-duration agent pricing; Growth starts at a $4,000 annual commitment for discounted rates and higher limits. Multichannel audio, add-ons, concurrency, auto-load settings, and model-improvement participation materially affect total cost.
Pricing notes that some displayed audio-intelligence rates opt in to the Model Improvement Program, making configuration and contract review important for sensitive data. Deepgram publishes SOC 2, HIPAA, GDPR/EU residency, CCPA, PCI, encryption, and role-control information. Those controls do not correct recognition errors, unsafe synthesized speech, or agent mistakes. Obtain consent, verify speakers and critical facts, restrict tools, and retain a human fallback.
How Deepgram works
Applications stream audio or submit recordings to a selected speech model, which returns text, timing, formatting, and optional speaker or language signals. Text-to-speech converts submitted characters into audio. The Voice Agent API coordinates listening, model reasoning, speaking, and tools, with bring-your-own model options. Humans must validate transcripts, voices, consent, tool actions, and model-specific billing.
How to set up Deepgram
Approve the speech data path
Map capture, consent, regions, model-improvement choice, retention, redaction, storage, voice rights, subprocessors, and human access.
Select models by workload
Benchmark prerecorded, streaming, conversational, multilingual, and TTS options on representative noise, accents, jargon, latency, and cost.
Harden the API path
Use scoped projects and keys, protected endpoints, authenticated callbacks, rate limits, safe file access, tenant isolation, and spend controls.
Validate optional intelligence
Measure diarization, terms, redaction, summaries, sentiment, intent, and turn detection separately rather than assuming one transcript score covers them.
Monitor production quality
Audit source audio and generated speech, track errors and latency, review auto-load, close streams, rotate keys, and escalate failed conversations.
Deepgram FAQs
Does Deepgram offer free usage?
New Pay As You Go accounts currently receive $200 in promotional credit before funded usage.
How is Deepgram speech-to-text priced?
Rates vary by model and streaming or prerecorded use and are billed by processed duration; add-ons and multichannel audio can increase cost.
Does Deepgram charge for silence?
Its pricing FAQ says speech-to-text uses per-second billing rather than rounding; streaming and agent session design should still be tested for idle cost.
Can Deepgram be self-hosted?
Enterprise customers can discuss containers for private-cloud or on-premises deployment, with hardware and contract requirements.
Are transcripts and summaries always accurate?
No. Audio quality, accents, names, jargon, speaker overlap, model choice, and downstream language processing all require representative tests and human review.
Listing reviewed 2026-07-24. Product details and pricing can change; verify important terms on the provider's website.
Related Automation AI tools
Related AI guides
Reviews
Tell the community what you made, what worked, and what you wish you knew before starting.